Pogbagram

Zhang Shuyu, born in 1999, is a product manager who recently joined Tencent's PC Manager team, which is not considered a core business line at Tencent.

In January of this year, OpenClaw became a sensation in China. She became fascinated with it and gathered a few people to create a product prototype, QClaw: based on OpenClaw, it can be installed with one click and the smart agent can be controlled directly through WeChat.

The project had almost no presence within Tencent's system; it lacked project approval and resources from the central office. It was just a group of young people getting together to write code.

On March 9th, QClaw launched its internal beta test. Within a week, millions of users registered.

Then things started to get out of control, alarming Tencent's general office.

The top management reacted swiftly, immediately allocating dozens of employees and computing resources to Zhang Shuyu's team. On the same day, another team launched WorkBuddy, also compatible with OpenClaw. The following day, Tencent's Hong Kong-listed shares surged by over 7%, with investors directly attributing the gains to these two companies.

At 2:06 AM on March 11, Ma Huateng posted on his WeChat Moments: "Self-developed lobster, local shrimp, cloud shrimp, enterprise shrimp, cloud desktop shrimp, secure isolation shrimp room, cloud security, knowledge base… and a batch of other products are coming soon."

This is a clear signal to Tencent's 110,000 employees, and countless employees interpret it as: Pony supports them going all in on lobster .

According to an exclusive report by The Information, as of this month, eight teams within Tencent are simultaneously developing products and services based on OpenClaw . Including projects under development and in internal testing, the total number exceeds ten.

Fifteen years ago, three teams within Tencent raced against each other to develop mobile instant messaging. Zhang Xiaolong's Guangzhou R&D department emerged victorious, creating WeChat – Tencent's most successful "horse race" in history. This time, the competition has shifted to a different species: shrimp.

It seems a bit unbelievable that a peripheral project done by a product manager born in 1999 could become a strategic pivot for a trillion-dollar company within two weeks.

Zhang Shuyu told The Information a very honest statement: "We are all experimenting with AI agents. At this moment, no one can say what the best method is."

In other words: We don't know the answer, but it's better to start running than to stand still.

The hope of the whole village: Why did Tencent stake its life on a shrimp?

To understand Tencent's passion for lobsters, one must first confront the company's current position in the AI ​​competition.

Over the past two years, China has been engaged in a fierce arms race in the field of large-scale AI models.

Alibaba has poured money into Qianwen, while ByteDance has incubated Doubao, giving them a significant lead in user scale and model capabilities. What about Tencent? While it enjoys substantial profits from games and WeChat advertising, its approach to AI is far less aggressive than its two rivals.

The self-developed Hunyuan large model is still unable to compete with its competitors, which has also hindered the progress of its own AI assistant "Yuanbao".

Tencent has not been idle. Last year, they brought in Yao Shunyu, a former OpenAI researcher, to lead the Hunyuan research team and rebuild their R&D infrastructure. The upcoming release of the next-generation Hunyuan model in April is widely regarded in the industry as a test of Tencent's modeling capabilities.

However, this distant solution cannot quench the immediate thirst. Before the new model is submitted, the lack of a strong internal model has temporarily put Yuanbao at a disadvantage in the competition with Doubao and Qianwen.

So when OpenClaw ignited the agent craze in China, Tencent's top management almost instinctively seized the opportunity. This lobster proves that the next breakthrough for AI may not be in the chat box, but rather on the desktop, in tools, and in countless intelligent agents that can do things for you.

Tencent's top management has a clear understanding: the wave of agents triggered by OpenClaw will be an opportunity to reshuffle the AI ​​battlefield .

Their logic is this: if Tencent can deeply integrate OpenClaw-like agent capabilities with WeChat and provide supporting tools and services to become the best agent usage platform in China, then even if its internal big model is not the most powerful and its AI assistant is not the most popular, Tencent still has the potential to turn the tide in the second half of the AI ​​era.

In 2020, Ma Huateng internally referred to Tencent's video platform as "the hope of the whole village," pinning his hopes on it to regain its footing in the short video arena. Now, "the hope of the whole village" has changed its species.

The difference is that the video platform is at least a product of the platform itself, while the lobster comes from the GitHub of an independent developer in Austria.

In a sense, this is more like what Nadella did after taking over Microsoft in 2014: admitting defeat in the mobile internet, letting go of his desire to "do everything himself" and betting on a completely new track.

Nadella took ten years; Tencent hopes to do it faster.

Eight shrimp vying for the throne, behind Tencent's shrimp competition

The outside world views the parallel operation of multiple teams as a classic horse racing mechanism, while Tencent internally prefers to call it "diversity." QClaw and WorkBuddy were the first two to emerge, taking completely different paths.

QClaw was developed by Zhang Shuyu, who emerged from the periphery of the PC Manager team. Embracing the OpenClaw open-source ecosystem, QClaw focused on one-click installation via WeChat and experienced rapid, unregulated growth. Its design philosophy can be summarized in four words: "Open and Use." No environment configuration or terminal command knowledge is required; simply scan with WeChat to let AI take over your computer.

WorkBuddy, on the other hand, took a completely different path. In an interview with APPSO, the person in charge, Wang Shengjie, repeatedly emphasized one thing: it's 100% self-developed, and they haven't used a single line of OpenClaw source code .

It takes a semi-automated approach, avoiding the risk of information being exposed on the public internet in OpenClaw's "transparent transmission" model. It uses a bot push notification model, requiring user confirmation for each critical operation. Wang Shengjie's definition is clear: Lobster is a concept, not equivalent to OpenClaw . WorkBuddy aims to provide a safe and controllable lobster, one that enterprises can use with confidence.

Wang Shengjie revealed a crucial timeline: WorkBuddy was launched on the weekend of January 17th, with three or four people working through the night to create an MVP (Minimum Viable Product), originally scheduled for release on March 16th. However, seeing the lobster craze, they moved it forward by a week, coinciding with the release of QClaw.

In other words, Tencent didn't rush to follow suit after OpenClaw became popular. Multiple teams sensed the same opportunity at different times, and OpenClaw's explosive popularity acted as a catalyst, pushing the previously unseen project to the forefront overnight.

However, the contradictions inherent in the shrimp-competition mechanism are also on the table.

QClaw and WorkBuddy have highly overlapping functions, and both can control AI agents via WeChat. Which one should users choose? With 8 teams running simultaneously, will resources be wasted?

The answer lies in Zhang Shuyu's words: "At this moment, no one knows what the best method is." With eight teams participating at the same time, it's less about overflowing confidence and more about no one being certain .

Tencent chooses to hedge against uncertainty with quantity, running multiple routes simultaneously, hoping to win just one.

The essence of a horse racing system has always been: increasing the probability of success through sheer numbers. That's how WeChat was developed 15 years ago.

Ma Huateng's shrimp farming philosophy

The premise of shrimp competition is that there are shrimp to compete with, but this shrimp is not under Tencent's control.

On March 12, Peter Steinberger, founder of OpenClaw, publicly criticized Tencent on X, pointing the finger at Tencent's SkillHub service for copying the community Skills without making any contributions.

Two days later, Tencent donated through GitHub and was subsequently listed as a featured sponsor, alongside OpenAI. At last week's NVIDIA GTC conference, Tencent Cloud CEO Tang Daosheng met with Steinberger in person, proposing that Tencent Cloud contribute server and security services, and to explore deeper cooperation with the OpenClaw Foundation.

A senior vice president of one of China's most valuable internet companies flew to San Jose to sit down with the founder of an open-source project to discuss cooperation. This is almost unprecedented in Tencent's history. When you need something from someone more urgently than they need something from you, you naturally become more humble.

At the earnings conference that same week, Tencent President Martin Lau announced that investment in new AI products would at least double by 2026, starting from 18 billion yuan last year. When explaining where the money would be spent, he only mentioned three products: Hunyuan, Yuanbao, and the latest Claw product .

Just a month ago, Lobster was a fringe project, but now it's on par with Tencent's self-developed large-scale models and flagship AI applications. Lobster has officially been upgraded from "something for everyone to play around with" to "company strategy" .

Ma Huateng's recent remarks at the earnings conference further answered a more fundamental question: What does Tencent want to do with lobsters ?

His approach skipped the product level and focused on the ecosystem.

Ma Huateng believes that lobster-themed apps have memory and personality, are more like assistants, and have a "living" feel, which allows AI to be implemented in various scenarios such as offices, terminals, and mini-programs, instead of all crowding onto the single path of chatbots.

But what's truly intriguing is his discourse on "decentralization." WeChat itself is a centralized app, but its ecosystem is decentralized, with hundreds of thousands of mini-program merchants forming an open platform. Ma Huateng believes that AI agents inherently possess decentralized characteristics and can integrate into the WeChat ecosystem. One sentence is particularly crucial:

All service providers are afraid of being "short-circuited" or "channelized" by AI intelligent agents.

This means he doesn't want the AI ​​Agent to become a new intermediary, turning service providers within WeChat into mere backend APIs. He wants mini-programs to retain their independence while possessing AI capabilities. " Every mini-program can be intelligent and 'lobster-like' in its own right. "

This thinking goes a step further than "we also make lobsters." Ma Huateng sees a possible paradigm shift: the way AI's value is distributed will change from "one super chatbot ruling everything" to "countless distributed intelligent agents each displaying their unique abilities."

If this assessment holds true, WeChat, with its world's largest communication ecosystem and most active mini-program platform, is naturally the most fertile ground for the Agent era .

At the earnings conference, Martin Lau clearly summarized this logic: "Claw proposed a decentralized model… For a time, it seemed that everyone was vying to become the sole entry point and monopolist for AI intelligent agents. But that's not the case."

In short, Tencent's betting logic can be summarized as follows: they lost the battle of models, but the battle of ecosystems has not yet been played out .

Of course, this narrative can also be translated into another sentence: Our model is not strong enough, so we're telling you that the model isn't that important.

The line between self-consistency and self-deception is sometimes very thin. But the key is that Tencent does have cards to play this time. WeChat doesn't need to be the container for the most powerful model; it only needs to be the most user-friendly agent runtime environment .

This is exactly the same logic as Nadella's Azure: you don't need to create the best AI yourself; you just need to let the best AI run on your cloud.

A panoramic view of shrimp farming products: How much has Tencent actually invested?

Tencent's "crab farming" strategy goes far beyond simply creating a few consumer-facing products. On Friday, Tencent released a "panoramic view of its crayfish farming products," a complete crayfish matrix from the bottom layer to the application layer, with a density exceeding external expectations.

Consumer-grade products are leading the charge. QClaw focuses on one-click installation via WeChat, targeting ordinary users; WorkBuddy takes a self-developed desktop approach, emphasizing security and controllability; and WeChat ClawBot allows users to control lobsters directly within the WeChat chat interface.

The three products cover three core scenarios: "easy onboarding for novice users," "deep desktop usage," and "seamless integration into the WeChat ecosystem." At the consumer level alone, Tencent has simultaneously laid out three paths.

Enterprise-level products followed closely behind. ClawPro targets enterprise and government clients, emphasizing secure isolation and granular permission control. It features an exclusive channel within WeChat Work, tiered account permissions, a built-in skills review mechanism, code generation operations requiring approval, and web searches using a security gateway.

At the Tencent Cloud Summit, Tang Daosheng highlighted ADP (Agent Development Platform), positioning it as a toolbox for enterprises to build customized agents. It works in conjunction with Claw Runtime to provide a secure sandbox environment, and Lighthouse for security management.

The logic behind the entire enterprise solution is very clear: OpenClaw is too wild, so I'll help you put it in a cage.

The developer ecosystem hasn't been neglected either. CodeBuddy, an AI programming assistant launched in the second half of last year, has now been incorporated into the Lobster Matrix as a developer portal; SkillHub is an AI skills community that has been localized, and it was precisely because this product was criticized by Steinberger that the subsequent donation was made. TokenHub is a model service marketplace, accepting not only Hunyuan but also third-party models such as DeepSeek, MiniMax, and Kimi, with unified billing.

Tencent has even come up with a business idea: "selling shovels."

This panoramic view shows that Tencent doesn't want to make a single breakthrough in its products; it wants to build an entire lobster industry chain— from installation to operation, from individuals to enterprises, from consumption to development, with someone monitoring every link.

This is precisely the "Harness Engineering" approach that Tang Daosheng repeatedly emphasizes: the key to success in the Agent era lies not in the model itself, but in the scaffolding. Tool usage, context engineering, long-term memory management, and workflow design—these seemingly unglamorous but arduous tasks are the crucial variables that determine whether an Agent is easy to use.

At the Tencent Cloud Shanghai Summit, Tang Daosheng stated, "The implementation of AI is not just a matter of algorithms; Harness engineering capabilities are a key variable. Different scaffolding designs will significantly affect the actual usage effect and token cost."

In layman's terms: the model is the engine, but without a chassis and steering wheel, it can't go very far. Tencent's model isn't as fast as others for now, but if they can perfect the chassis and steering wheel, they can still win.

After the shrimp tide recedes

If you string all the clues together, the story can be condensed into one sentence: Tencent used all the resources that a large company can mobilize to embrace an open-source project that it could not control .

This is a posture full of tension.

OpenClaw updates two or three versions per week, its API is changed on a whim, and breaking changes come at any time. Peter clicks "merge," and several product teams in a Shenzhen building might have to work through the night to put out fires. Tencent's strategic lifeline is tied to someone else's GitHub repository; this requires not only courage but also an unprecedented level of humility.

But from another perspective, Tencent probably didn't have a better option.

If we continue to compete head-on only in the models and chatbot race, we'll either be mere runners-up or get bogged down in a homogeneous battle. But the agent wave has opened up a new niche: whoever can turn AI into the most user-friendly tool will be able to redefine the entry point .

WeChat boasts 1.4 billion monthly active users, a mini-program ecosystem, payment capabilities, and a social relationship network. These elements alone cannot create the strongest model, but they can create the best agent usage environment—a unique advantage Tencent holds that no one else possesses.

The question is, how long is this card valid?

OpenClaw is still rapidly iterating, and its ecosystem is far from finalized. Will today's lobster craze be as fleeting as last year's Manus? Will the eight teams competing in the lobster competition produce the next WeChat, or just eight half-finished products? Ma Huateng's blueprint for a "decentralized agent ecosystem" is beautiful, but how many more "technical mishaps" will it take to go from blueprint to reality?

However, one thing is certain.

When a company's CEO posts on WeChat Moments at 2 a.m., the president lists lobsters and self-developed models side by side in an earnings call, the senior vice president flies to the United States to meet with the founder of an open-source project, eight teams compete in a lobster race, and AI investment doubles, it's no longer just chasing trends; it's betting on the company's future.

The bet isn't about how long this shrimp can live . The bet is about whether Tencent can still be at the table, and in what position, during the decade when AI reshapes everything.

WeChat Channels was once hailed as "the hope of the whole village." Five years later, it hasn't defeated Douyin, but it has carved out its own niche within the WeChat ecosystem. Can lobsters also find a third way? It's too early to say.

However, when a giant is cornered and finally figures out what it wants, and pours resources into a single direction, you should never underestimate it.

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.

Apple's AI feature was unexpectedly launched in China in the early hours of the morning but was subsequently removed.

Early this morning, Apple Intelligence suddenly appeared in the settings page of Chinese iPhone users, but Apple quickly removed it. Bloomberg reporter Mark Gurman stated explicitly that the feature has not yet been officially approved, its appearance was accidental, and it is unrelated to the iOS 26.5 beta.

APPSO discovered that the original "Siri" entry in the settings page of Chinese iPhones has been renamed "Apple Intelligence & Siri." Clicking it allows users to download on-device models and unlock a range of AI features. With only one day left until Apple's 50th anniversary (April 1st), Apple Intelligence has quietly arrived on the phones of Chinese iPhone users.

Related Reading: Breaking News! Apple's AI feature was released in China ahead of schedule; after overnight testing, we discovered these details.

Gurman provided three reasons for judging that this was an accident:

Apple will not launch AI features in China's most important market without any official announcement; Apple will not choose to release them in the early hours of the morning in China; in addition, the feature calls Google's visual recognition engine, and Google services are not normally accessible in China.

APPSO activated and tested the features immediately before they were taken offline. The features launched include a new Siri interface, writing tools, a picture gallery, AI-powered image removal, visual intelligence, and translation.

Regarding model invocation, the situation is more complex. The visual recognition engine, according to testing, comes from Google; in the Siri dialogue and content generation stages, APPSO's testing revealed instances of GPT being invoked, and some users online have also reported invoking Baidu's Wenxin large model.

This is significantly different from the industry's previous general expectation that "the Chinese version will only access Baidu and Alibaba models".

Apple Intelligence made its debut at WWDC24 in June 2024 and was first launched in the US in October of the same year with iOS 18.1. Last March, the iOS 18.4 update enabled it to support Simplified Chinese.

However, the release date of the Chinese version has been postponed several times since then, from the initial target of mid-last year, to iOS 18.6, iOS 26.1, iOS 26.2, and iOS 26.4.

In November of last year, Gurman stated in his Power On column that the release of the Chinese version was "a long way off," and pointed out that in addition to approval issues, the engineering progress of Apple Intelligence itself was also not going smoothly, and the model performance did not meet expectations.

Apple has since removed the unexpectedly released version, and the official release date for the Chinese version of Apple Intelligence remains unclear.

Zhang Xue's motorcycle wins WSBK Portugal round for the second consecutive time, breaking the monopoly of Europe, America, and Japan for the first time.

According to reports from Jimu News and Jiemian News, from March 28th to 29th local time, Chinese motorcycle brand ZXMOTO won the championship in the SSP middleweight category at the WSBK Portugal round for two consecutive rounds, becoming the first Chinese motorcycle brand to win a WSBK round, breaking the decades-long monopoly of European, American and Japanese manufacturers such as Ducati and Yamaha.

Related reading: China's first championship! Taking on Ducati and surpassing Yamaha, Zhang Xue has become the "Lei Jun" of the motorcycle world.

French driver Valentin Debis achieved his second consecutive championship driving the brand's self-developed 820RR-RS race car.

The motorcycle is equipped with a self-developed three-cylinder engine with a maximum speed of 16,000 rpm. It is worth noting that Zhang Xue's motorcycle only achieved 19th and 24th place in the first leg of the Australian Grand Prix in February this year, so this victory can be described as a complete reversal in just over a month.

The brand's founder, Zhang Xue, was born in 1987 in a rural area of ​​Huaihua, Hunan Province. He started as a car repair apprentice and only founded Zhang Xue Motorcycle in April 2024, less than two years ago.

In addition, after actor Yin Zheng posted a congratulatory message on Weibo, Zhang Xue responded that she "didn't have the money" to hire a spokesperson and decided to directly gift him an 820, asking him to share his honest experience after riding it. The standard version of the 820RR is priced at 43,800 yuan, and the 820RR-R version is priced at approximately 61,000 yuan.

Apple's open-source programming language Swift officially supports Android.

Recently, Apple's open-source programming language Swift officially launched support for the Android platform, and the first Swift SDK for Android has been officially released, nearly a year after the plan was officially announced.

With this SDK, developers can start developing native Android applications in Swift, update existing Swift projects to support Android builds… This is a significant milestone that opens up new opportunities for cross-platform development in Swift.

From a technical implementation perspective, building a Swift package on Android requires cross-compilation, which means compiling the code into a program that can run on Android devices or emulators on desktop platforms such as macOS or Linux.

The complete toolchain consists of three parts: the Swift toolchain (the core compiler), the Swift SDK for Android (an extended support package specifically for Android cross-compilation), and the Android NDK (which provides platform-specific header files, system libraries, and linker tools). Currently, this SDK supports both x86_64 and aarch64 Android architectures.

In theory, a single Swift codebase could potentially run on both iOS and Android platforms, significantly reducing the barriers and costs of cross-platform porting. Of course, Kotlin remains the dominant language for Android development, and the addition of Swift primarily expands developers' options rather than replacing the existing ecosystem.

Details of OpenAI shutting down Sora: Millions of dollars in daily losses; Disney executives only learned of the shutdown an hour before the official announcement.

Yesterday, The Wall Street Journal revealed details of OpenAI's recent shutdown of its video generation app, Sora. This product, which CEO Sam Altman likened to a "GPT moment," went from its public launch last September to its complete closure in less than six months.

The report points out that the shutdown decision directly impacted the partnership between OpenAI and Disney. Last December, the two companies announced a multi-year cooperation agreement, with Disney committing to licensing more than 200 IP characters and investing $1 billion.

However, several Disney executives only learned of the news less than an hour before the official announcement. Currently, the investment has never materialized, and the collaboration has effectively stalled. Disney's new CEO, Josh D'Amaro, is in discussions with more than ten partners regarding the introduction of other AI tools.

The report also stated that Altman initially provided Sora with substantial resources because he hoped OpenAI would become a company that uses AI to reshape popular culture and entertainment.

In an internal letter to employees, Altman characterized the shutdown as a "difficult but necessary choice," and stated that the Sora team would shift its focus to longer-term areas such as robotics. Furthermore, OpenAI is pushing forward with its new model codenamed "Spud" and plans to create a "super app" primarily for productivity scenarios.

In its early days, Sora topped the App Store charts and had a peak of about 1 million users worldwide, but subsequently dwindled to less than 500,000, losing about $1 million a day.

The training cost of video models is much higher than that of language models. Given the increasingly strained computing resources at OpenAI, Sora is internally regarded as a burden that is "not worth the effort".

ClawBench's latest rankings: Zhipu GLM-5-Turbo tops the list, while four models from ByteDance and Xiaomi enter the global top ten.

Agent benchmarking organization ClawBench released its latest large-scale model ranking list yesterday, covering 30 complex agent tasks and encompassing five core business scenarios: office collaboration, information retrieval, content creation, data processing, and software engineering.

This list includes over 40 mainstream large models, with four domestic models from Zhipu, ByteDance, and Xiaomi making it into the global top ten.

• The Zhipu GLM-5-Turbo topped the list with a CLAW SCORE score of 93.9, becoming the best-performing model in this evaluation.
• ByteDance's Doubao-Seed-2.0-lite ranked second with a score of 93.1, and its usage cost was only $0.33, the lowest on the entire list;
• The Xiaomi MiMo-V2-Omni ranked 9th with a score of 91.2, and its running speed was the fastest on the entire list, completing the entire task process in just 848 seconds.

Looking at the overall rankings, OpenAI GPT-5.4 ranked third with a score of 92.2, Claude Opus 4.5 ranked seventh with a score of 91.5, and Alibaba Qwen3.5-35B-A3B ranked eighth with a score of 91.4.

ClawBench employs an isolated sandbox execution mechanism, requiring each model to complete tasks in a realistically simulated enterprise development environment, and deliberately incorporates engineering challenges such as "inconsistent naming," "missing directories," and "date traps."

In terms of scoring, ClawBench introduces a "triple scoring mechanism," which uses automated script assertions, cutting-edge LLMs as "expert judges," and a weighted combination of both, depending on the task type, in order to more realistically reflect the model's actual deployment capabilities in complex workflows.

iMac models equipped with OLED displays may enter mass production between 2029 and 2030.

According to ZDNet Korea, Apple has recently requested Samsung and LG, two OLED panel manufacturers, to use their mass production lines to produce OLED samples for the iMac, in preparation for the OLED version of the iMac expected to launch in 2029-2030.

It is understood that Apple proposed a 24-inch OLED screen for the iMac to both manufacturers, with a brightness of 600 nits and a pixel density of 218 PPI, which is an improvement over the 500 nits of the current LCD version of the iMac, while the pixel density remains the same.

Lei Jun personally announced the recruitment of AI talent.

Lei Jun, founder, chairman and CEO of Xiaomi, announced yesterday that Xiaomi's AI talent recruitment program has officially started.

Lei Jun stated that Xiaomi's R&D and capital expenditure in the AI ​​field this year reached 16 billion yuan, and significant progress has been made in projects such as large-scale foundation models and embodied intelligent robots. This recruitment program features three channels: top talent, campus recruitment, and internships, creating a "one-stop, borderless" AI talent introduction system.

Our top talent recruitment is open to recent graduates from 2024 to 2027, postdoctoral researchers completing their research in 2027, and undergraduate, master's, and doctoral students graduating in 2027 and beyond. Positions are available across multiple fields, including large-scale models, mobile phones, automobiles, and autonomous driving.

Lei Jun explained that AI technology is fully empowering Xiaomi's entire ecosystem of people, cars, and homes, covering areas such as smartphones, smart cars, wearable devices, smart homes, and robots.

As previously reported, the MiMo model was developed by Xiaomi's core technology team, Core Team. The team's average age is only 25, with more than 60% of the members being graduates of Tsinghua and Peking Universities, and 55% holding doctoral degrees. The youngest core researcher is a 19-year-old university intern.

A memory sell-off is underway in Shenzhen's Huaqiangbei district: 32GB DDR5 memory is now priced as low as 1950 yuan.

According to CLS News Agency, a reporter visited the Huaqiangbei Electronics Market in Shenzhen yesterday and learned that the spot price of DDR5 memory modules has dropped significantly since last week.

Among them, the price of a single 32G DDR5 memory stick, which was priced at about 3,000 yuan last week, has dropped by 500 to 1,050 yuan this week, with some merchants even selling it for 1,950 yuan per stick .

One merchant stated, "I just cleared out a batch of products at 2,500 yuan per unit," and some DDR5 products are currently being sold off. Opinions among the merchants interviewed regarding future market trends differ: some believe prices still have room to fall further, while others believe upstream prices are providing support, and this round of price reductions may be unsustainable.

An executive from a listed company with a memory module business analyzed that this round of selling may be related to weak market demand for short-term installations and downstream merchants' desire for rapid turnover. However, he emphasized that the price correction will not affect the upward trend of the entire storage industry, including memory modules.

One month after its launch, Kimi K2.5's Dark Side of the Moon ARR surpassed $100 million.

According to Jiemian News, about a month after the launch of Kimi K2.5, the Dark Side of the Moon's ARR (Annual Recurring Revenue) surpassed $100 million in early March.

Sources familiar with the matter revealed that after the launch of K2.5, the TPM (tokens per minute) quota for API supply quickly became tight, with some clients offering consumption commitments and prepayment guarantees in the tens of millions of dollars in order to obtain priority supply qualifications.

Following its $500 million Series C funding round at the end of last year, Dark Side of the Moon was reported to be nearing completion of a new funding round of over $700 million in mid-February this year. According to the Science and Technology Innovation Board Daily, its latest valuation has risen to $18 billion, quadrupling in three months, and it is currently undergoing a $1 billion funding round.

In addition, last week media reports indicated that Dark Side of the Moon was considering an IPO in Hong Kong.

Oriental Selection's first store will open next month in Beijing's Zhongguancun. Yu Minhong: We plan to open hundreds of offline supermarkets in the future.

Recently, Yu Minhong, Chairman and CEO of Oriental Selection, revealed in a live broadcast that Oriental Selection's first offline physical store will officially open in Zhongguancun, Beijing in April. The store covers an area of ​​approximately 400 square meters and offers a range of convenience store products, including fresh produce, snacks, and daily necessities, as well as a light meal and coffee beverage area.

Yu Minhong stated that once the first-store model matures, it will leverage New Oriental's approximately 80,000 employees across the country to establish two to three Oriental Selection supermarkets in each city, at which point the total number of stores nationwide will reach dozens or even hundreds.

In terms of product structure, Oriental Selection's self-operated products account for about one-third, while the rest are high-quality products from all over the country.

Paradigm Intelligence achieved its first full-year profit, with annual revenue of 7.135 billion yuan.

Yesterday, Paradigm Intelligence released its full-year 2025 results announcement.

Total revenue for the year was RMB 7.135 billion, representing a year-on-year increase of 35.6%, with a gross profit margin of 34.8%. Adjusted net profit attributable to shareholders was RMB 17.84 million, achieving a turnaround from loss to profit. Calculated under International Financial Reporting Standards, the annual loss attributable to owners of the parent company was RMB 26.3 million, a significant reduction of 90.2% year-on-year.

The performance of the three major business segments is as follows:

• AI Platform : Revenue of RMB 6.552 billion, up 32% year-on-year, accounting for 91.8% of total revenue, is the core growth pillar, benefiting from the demand for domestic substitution and the "AI+" policy.
• API business : Revenue of RMB 79.9 million, a year-on-year increase of 129.2%, the fastest growth rate; Token revenue in the first quarter of this year has already exceeded the total for the whole of last year;
• Agentic AI business : Revenue of RMB 503 million, up 93.2% year-on-year, focusing on the energy, finance and other industries with the "pay-for-results" model.

SF Express 2025 Annual Report: Net profit attributable to parent company was 11.1 billion yuan, with a slight narrowing of gross profit margin.

Yesterday, SF Holding announced its audited results for the year ended December 31, 2025.

Full-year operating revenue was RMB308.23 billion, an increase of 8.4% year-on-year, exceeding RMB300 billion for the first time; profit attributable to owners of the company (net profit attributable to the parent company) was RMB11.12 billion, an increase of 9.3% year-on-year; basic earnings per share were RMB2.23.

The core data for each business segment are as follows:

• Time-sensitive express delivery: Revenue reached RMB 131.05 billion, a year-on-year increase of 7.2%;
• Economy Express: Revenue of RMB 32.05 billion, up 17.6% year-on-year;
• Express delivery: Revenue of RMB 42.13 billion, up 11.9% year-on-year;
• Same-city instant delivery: Revenue reached RMB 12.87 billion, up 42.8% year-on-year, with net profit more than doubling to RMB 280 million, a record high;
• Supply chain and international business: Revenue of RMB 76.35 billion, up 3.2% year-on-year, with the segment turning a profit (net profit of approximately RMB 190 million).
• The total number of parcels handled throughout the year exceeded 16.7 billion, representing a year-on-year increase of 25.4%. The gross profit margin was 13.07%, a decrease of 0.61 percentage points from the previous year, mainly due to the increase in the proportion of labor costs.

Google is reportedly developing a "tap to transfer" file feature for Android.

According to Android Authority, Google will introduce a feature called "Tap to share" for Android's built-in Quick Share, similar to Apple's "tap to share" file transfer feature in the new version of AirDrop.

The earliest clues appeared last September. In the experimental feature area of ​​Samsung One UI 8.5, a prototype of NFC-based file sharing appeared, but at the time, the feature seemed more like an internal Samsung test project.

In the leaked version of One UI 9 that surfaced yesterday, this feature has been explicitly named "Tap to share," and its operation logic is straightforward: simply bring the tops of two phones close together to send files.

In this year's Android 17 beta and Canary versions, a system-level service called "TapToShare" appeared. This service relies on Google Play services to run, which means that this feature is expected to become a general Android capability rather than a feature exclusive to a particular brand.

iQiyi plans a secondary listing in Hong Kong.

According to a report by Yicai Global, iQiyi announced on March 30 that it had confidentially submitted a listing application to the Hong Kong Stock Exchange, applying for permission to list and trade its Class A ordinary shares on the Hong Kong Stock Exchange.

On the same day, iQiyi's board of directors approved a share repurchase program, authorizing the company to repurchase up to $100 million worth of shares (including in the form of American Depositary Shares) over the next 18 months, effective immediately from the date of approval.

Li Auto's Mach 100 chip paper has been shortlisted in top-tier chip academic circles.

Yesterday, Li Auto officially announced that its research paper on the Mach 100 chip, titled "M100: An Orchestrated Dataflow Architecture Powering General AI Computing," has been accepted by ISCA 2026 (International Conference on Computer Architecture). Li Auto has thus become the first company in the automotive industry to have a paper accepted in the conference's Industry Track.

ISCA is one of the most authoritative and top-tier academic conferences in the field of computer architecture. It has long focused on fundamental technical issues such as chips, processors, and AI computing power, and is regarded as a trendsetter in the field.

The paper introduces the "meticulously orchestrated data flow architecture" used in the Mach 100 chip that is about to be mass-produced and put into vehicles. The core idea is to allow data to flow directly between computing units, minimizing the repeated access to the cache.

Li Auto stated that this architecture design makes the Mach 100 chip more efficient in data processing, enabling it to release more effective computing power, while also possessing stronger programmability to adapt to the rapid iteration of AI.

Bloomberg: AI is stealing the first jobs from young people in London.

According to Bloomberg, London is experiencing a structural crisis in youth employment driven by AI, and the impact of this crisis is disproportionately concentrated on young people just entering the workforce.

According to data cited by Bloomberg, about one-third of London’s workforce works in industries highly exposed to the risks of being replaced by AI, such as professional services, administration, IT, and finance. This proportion is significantly higher than the UK average of about one-quarter.

Daniel Harris, managing director of Robert Walters UK and Ireland, points out that white-collar employers are using automation on a large scale to replace entry-level positions or are shifting recruitment to regions with lower labor costs.

The report also points out that the impact of AI is asymmetrical. It tends to reward experienced professionals while prioritizing the elimination of entry-level positions where employees haven't yet accumulated work experience. This means that recent graduates, who most need their first job to gain experience, are precisely the group most likely to overcome this hurdle.

According to statistics from the recruitment website Adzuna, the number of job openings for recent graduates in London has plummeted from approximately 13,000 in 2019 to approximately 2,000 at the beginning of this year, a drop of more than 85%. The proportion of London job openings in the total number of recent graduate job openings in the UK has also shrunk from one-third a decade ago to one-fifth today.

Meanwhile, the unemployment rate for 16- to 24-year-olds in London has climbed to 25%, the highest in the UK and higher than that of their peers in major European cities such as Madrid and Paris.

A recent computer science graduate frankly admitted, "Everyone is talking about AI replacing software developers right now, and it's really too difficult to compete with a technologist who makes mistakes but can be easily corrected by senior engineers."

Starting at 6999 yuan, the vivo X300 Ultra has been officially released.

Last night, vivo launched two new smartphones at its spring product launch event: the X300 Ultra and the X300s. The former is positioned as a professional imaging flagship, while the latter is a flagship with balanced imaging performance.

vivo X300 Ultra:

• Equipped with the fifth-generation Snapdragon 8 Ultra;
• The 3+2 Zeiss Master Lens lineup includes a 50MP 14mm Sony LYTIA 818 ultra-wide-angle lens, a 200MP 35mm Sony LYTIA 901 street lens, a 200MP Samsung HP0 85mm periscope telephoto lens, and a Blueprint camera, supporting full-range 4K 120fps Dolby Vision output and 4K 120fps 10-bit Log.
• Optional 200mm and 400mm teleconverters are available;
• Four-microphone array; supports satellite communication (limited to 16GB+1TB version);
• Available in three colors: "Film Green", "Black Ka", and "Silver".
• Prices start at 6999 yuan (12GB + 256GB), the 16GB + 1TB satellite communication version is priced at 8999 yuan, and the photographer's kit is priced at 11999 yuan;

vivo X300s:

• Powered by MediaTek Dimensity 9500;
• 200MP Samsung HPB main camera, Sony LYTIA 602 telephoto lens;
• 6.78-inch flagship large flat screen, 8T LTPO + Q10 Plus luminescent substrate, 144Hz high refresh rate, peak brightness of 2000 nits;
• Available in four colors: titanium black, silver white, film green, and dream core purple;
• Prices start at 4999 yuan, the 16GB + 1TB photographer kit is 7999 yuan, and the G2 teleconverter is priced at 999 yuan.

Related reading: vivo X300 Ultra Image Review: An Imaging Phone That Finally Becomes a Camera

Qwen3.5-Omni launched: 215 state-of-the-art features, plus the ability to "listen to videos and write code".

Yesterday, Qianwen released its latest generation of full-modal large model, Qwen3.5-Omni, which supports the understanding and generation of text, images, audio, and audio-visual content.

The model adopts the Thinker-Talker architecture and provides three specifications: Plus, Flash, and Light. It supports a 256k long context and can handle more than 10 hours of audio input and more than 400 seconds of 720P audio and video input. The training data covers more than 100 million hours of audio and video content.

Compared to the previous generation Qwen3-Omni, the key upgrades include:

• Multilingual capabilities : Speech recognition supports 113 languages ​​and dialects, and speech generation covers 36 languages ​​and dialects;
• Audio and video understanding : The Plus version achieved state-of-the-art (SOTA) scores in 215 sub-tasks/benchmarks across audio/audio and video related tasks, with overall capabilities reaching the level of the Gemini 2.5 Pro;
• Audio/video caption : Can generate structured, script-level fine-grained descriptions with timestamps;
• Audio-Visual Vibe Coding : The model can directly generate code based on audio and video instructions.

In terms of real-time interaction, the new version natively supports semantic interruption, WebSearch and complex Function Call calls, end-to-end voice control, and voice cloning.

OPPO Pad mini revealed

Yesterday, blogger "Digital Chat Station" posted on Weibo about the appearance and related configuration information of the OPPO Pad mini, and the new product is expected to be released in April.

• It is equipped with a Qualcomm Snapdragon 8 Gen 5 processor, and comes standard with 16GB of RAM and 512GB of storage;
• It uses an OLED screen with extremely narrow bezels on all four sides;
• Supports 67W wired fast charging;
• It is available in three color schemes: dark gray, purple, and cyan.

WeChat Work's open-source CLI allows AI to directly access office capabilities such as messages, calendars, and documents.

On March 30, WeChat Work launched the CLI open-source project on GitHub, opening up seven core product capabilities including messaging, calendar, documents, smart tables, meetings, to-do lists, and contacts, and supporting calls from mainstream AI agents such as Claude Code, Codex, WorkBuddy, and QClaw.

For installation and use, developers can install the CLI via npm, configure the Bot ID and Secret of the WeChat Work robot, and then start calling the relevant capabilities. The current project already provides a series of skills including messaging, meetings, to-do lists, calendars, documents, and smart spreadsheets.

Pop Mart officially announced its first crossover home appliance product, the LABUBU refrigerator.

Recently, Pop Mart officially announced its first product in the home appliance industry, the LABUBU refrigerator, which belongs to THE MONSTERS lifestyle series.

Based on the official images, this refrigerator features a compact and exquisite design. The doors are printed with LABUBU and TYCOCO, classic IP characters from Pop Mart, and the door handles feature a 3D embossed LABUBU design. Currently, the official price, specifications, and release date have not yet been announced.

Previously, Wang Ning, CEO and Chairman of Pop Mart, announced at the 2025 performance release conference that the company will officially launch a series of derivative small home appliance products based on its own IP in April. The first batch will be sold simultaneously on mainstream e-commerce platforms such as JD.com. The market strategy is to take a step-by-step approach of "first domestic, then overseas".

According to Blue Whale Technology, citing sources familiar with the matter, Xinbao Co., Ltd. is the core OEM manufacturer for Pop Mart's small appliance business, and the two companies have a deep cooperation under the OEM model. Currently, the product categories cover multiple items including electric kettles, coffee machines, electric toothbrushes, and hair dryers, and the company is now in the stage of large-scale inventory preparation.

Doubao is reportedly testing an "AI shopping" feature.

According to the Science and Technology Innovation Board Daily, Doubao has integrated with Douyin e-commerce, allowing users to order and pay directly within the Doubao app without being redirected to Douyin. This feature is currently in internal testing.

According to Bianews, after entering a question containing keywords such as "recommend a product" in the Doubao app, Doubao will not only recommend products but also pop up product links. Furthermore, users can view their orders placed through Doubao in the "My – Settings – My Orders" section of the Doubao app.

Luckin Coffee's "coconut egg" has been criticized for being too cold, having a small portion, and not being sealed; stores are limited to 10 per order.

According to Jiupai News and China News Service, Luckin Coffee's recently launched "Coconut Egg" (a whole raw coconut) has sparked a lot of complaints from consumers on social media platforms.

Multiple consumers reported several issues with the product: the drinks were extremely cold, almost entirely in a slushy state, with staff even suggesting customers wait half an hour before drinking; the portion size was only half that of similar products; the coconut meat had a poor texture; and the product was not sealed, making it prone to spillage during takeout and delivery.

In response, Luckin Coffee customer service stated that they have received feedback from relevant users. Currently, the product supports ordering with ice removed. Issues regarding portion size, taste, and packaging will be recorded and reported to the relevant departments for processing.

Meanwhile, the product is experiencing supply shortages in stores across multiple districts of Beijing. Stores in Xicheng, Haidian, Shijingshan, and other districts of Beijing have indicated that sales are temporarily suspended.

Two-wheeled electric vehicle brands may collectively raise prices next month.

According to Blue Whale News, the domestic electric two-wheeler market may see a wave of concentrated price increases starting in April. Leading brands such as Ninebot, Yadea, Tailg, and Aima plan to raise prices on most of their models, with the increase expected to be between 200 and 300 yuan.

The report points out that regional managers of various brands have already intensively notified distributors to increase their inventory levels in order to cope with the potential increase in inventory costs and short-term demand fluctuations after the price increase.

According to CBN, Ninebot confirmed yesterday that due to rising raw material prices, the suggested retail price discounts for some models will be reduced by 100 yuan starting in April.

The first trailer for "PAW Patrol: Dinosaur Island" has been released, with the Backstreet Boys contributing the theme song.

According to The Hollywood Reporter, Paramount Pictures' animated film "PAW Patrol: The Rise of Dinosaurs" officially released its first teaser trailer and first promotional poster yesterday, confirming its North American release on August 14 this year.

This film is the third installment in the Paw Patrol series, directed by Carl Brunkel. The story follows the Paw Patrol dogs who are unexpectedly crash-landed on a tropical island inhabited by dinosaurs after a mysterious storm. In order to stop the villainous "Bad Mayor" Hardinger from mining wildly and triggering a volcanic eruption, they embark on a series of thrilling "dinosaur-level" rescue missions.

Notably, the Backstreet Boys recorded a brand new single, "Bottle Up," specifically for this film, which will be released as the film's theme song.

HBO has confirmed that the second season of the Harry Potter series has entered development.

Even before the first season of HBO's Harry Potter series has aired, development of the second season has officially begun. HBO executive Casey Bloys told Variety that the production team is actively working to avoid excessively long gaps between seasons.

Bloys stated, "Our goal is to avoid large gaps, especially considering the children are growing up." He also admitted that due to the show's large scale, it's impossible for it to air a season every year, but "they're already writing the second season."

The first season of *Harry Potter and the Sorcerer's Stone* is scheduled to premiere on December 25th this year. According to HBO's overall plan, the series will have seven seasons, one for each book, with eight episodes per season. The second season is expected to be an adaptation of *Harry Potter and the Chamber of Secrets*.

It is worth noting that the trailer for the series quickly attracted widespread attention after its release, reportedly accumulating over 277 million views within 48 hours of its release, becoming the most-watched trailer in HBO history.

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.

I typed in a prompt and pressed generate. My hands were still on the keyboard when the video appeared.

The foreground of the image is a steaming cup of black coffee, with a blurred background; then the focus smoothly shifts, clearly showing a girl flipping through an old book in the background.

I changed two words, adjusted the light direction, and pressed it again. It reappeared, and the texture wasn't lost. The prompts were written casually; it was just a plain description of the image in my mind, without deliberately piling on keywords, and the model understood it all.

This is my first impression of the PixVerse V6. It's not just about "good results"; to be precise, it's about "the effect, and the speed at which it came out."

PixVerse's fast video generation speed is hardly news in the industry. While most AI video tools are still testing patience with long queues, PixVerse is already the fastest in its class. In Artificial Analysis's video generation model leaderboard, PixVerse V6 is even in the top tier.

However, if V6 is just "a little faster and with better graphics", then it's just a regular iteration and not worth discussing separately.

What V6 does is transform "speed" from a technical parameter into a structural advantage at the creative level. Improved command comprehension means fewer repetitive tweaks to prompts; image quality is also more stable. The time saved in these areas, when combined, is far more valuable than a simple increase in generation speed.

In other words, when the quality of output is consistent and the speed is fast enough, the bottleneck for creators is no longer the tools, but the ideas themselves. AI has not devalued creators; on the contrary, it has made good ideas more valuable than before.

Not only is it fast, but every frame can withstand close scrutiny.

PixVerse V6 is easy to use and even beginners can easily get started.

Open Paiwo AI (web version: https://pai.video), describe the desired scene in your own words in the input box, select the resolution, aspect ratio, and duration, and then click "Generate". It supports output from 360P to 1080P, and can be used in both landscape and portrait modes. You can control the duration of each generation.

V6 has made visible progress in understanding instructions. Simply describe your creative intentions; there's no need to pile up technical jargon. It understands what kind of lighting, atmosphere, and shot you want. This saves not only time but also the energy of repeatedly refining prompts.

It also supports uploading reference images. If you want to recreate a certain style or lock in the character's appearance, just put in the image, and the model will be generated accordingly.

The most surprising improvement in V6 for me is in physics simulation.

I tested the macro lens. Golden honey dripped slowly, forming glistening threads that spread gently across the muffin. The threads were thinned by gravity, and the rate of diffusion on the contact surface matched the viscosity of the honey.

It has weight, gravity, and cause and effect.

Previously, the physical world in AI videos was "moving pixels," but in V6, objects obey the laws of physics. This has significance beyond the visual level. We judge whether an image is real or not based on intuition: Is the way this glass of water was poured correct? Is the trajectory of this drop of honey reasonable? V6 delivers a decent answer on this level.

Physical simulation addresses whether the world is realistic, while human portrait texture addresses whether the texture of a person is accurate.

This is the subject where AI videos are most prone to mishaps. A slight misstep can lead to the uncanny valley. The features are delicate but the expression is stiff, the skin is smooth but feels like silicone, the eyes are bright but lack focus. You can recognize it as a face, but your intuition will tell you it's not a person.

V6 has given me a fresh perspective on this aspect. Some creators in the industry compare it to Seedance, and the general conclusion is that both have their strengths and weaknesses, and it's hard to say which one is superior.

I tried a close-up of a middle-aged male actor's face, as he struggled to suppress the urge to cry, using shallow depth of field and soft side lighting. The result was captivating. The slight trembling of his nostrils, the moistness in his eyes—these movements were timed, layered, and logically connected.

This subtle nuance in micro-expressions gives the character a sense of being an AI actor. In contrast, previous AI portraits conveyed emotions more like "state switching," while V6's emotions show a more nuanced progression. The skin texture is also noteworthy—texture, pores, fine lines, and color variations under different lighting are all preserved.

The fight scenes are a physical fitness test for AI videos.

The punch must land in the right spot, the person being punched must react appropriately, the body's movement momentum must be continuous, and the spatial relationships between multiple characters must not clip through each other. If any of these elements break down, it becomes a comedy video. Many AI video models generate results that either show two people doing gymnastics or a punching arm passing right through the other person's face.

During my experience with V6, I generated a scene of two martial arts masters engaged in intense close combat in a muddy alley during a torrential downpour. A powerful punch landed on the opponent's jaw, rain and sweat splashed with the impact, and the victim's facial muscles trembled realistically from the force of the impact.

Beyond image quality, what truly excites me is that the V6 completes a set of visual language capabilities.

In terms of transformation effects, I generated a highly technological mechanical drone that flew rapidly in the air, and then quickly transformed from a metallic form into a water-ink dragon composed of flowing ink and ink lines, emitting a ghostly blue light.

The transition is natural, without the common pixelation and edge ghosting. There is a transitional state between the hardness of metal and the fluidity of ink; neither texture is abruptly cut off. The entire transformation process is closer to fusion than replacement.

The camera movement capabilities are equally impressive. From a cinematic aerial perspective, a drone rapidly traverses the interior of an abandoned, rusty industrial factory. The sense of space within the factory, the handling of light to create a metallic feel, and the stability of the footage during high-speed camera movements are all executed flawlessly, without exhibiting the chaotic feeling of "not knowing where to look" often seen in AI-generated footage.

Another approach is bullet time. It's a Matrix-style perspective rotation where the subject freezes while the background continues in motion. This type of shot demands a high degree of control over the sense of time; even a slight deviation can result in slow motion. The V6 handles this with restraint, showing some speed gradation without overdoing it.

I also tried two scenes that required even greater spatial depth. The first was a classroom. The drone's perspective cut from the corridor into the classroom, and the camera followed the airflow out of the classroom, zooming out to cover the entire campus. Exam papers and blank sheets of paper were swirled up by the airflow, filling the sky, but the camera movement remained continuous.

The second scenario involves bees squeezing in through window cracks, passing through the bedroom and living room, finding the honey jar in the kitchen, and then flying out again. Each room has different lighting and depth of field, and the bees' flight inertia is slightly delayed during scene transitions, mirroring the rhythm of real insects.

Multi-camera cuts are perhaps the most significant contribution to actual workflows. A short fashion advertisement for women's clothing features three shots seamlessly connected, with a warm and unified color palette, avoiding the patchwork feel of AI editing.

There are also three tests that I particularly like: camera movement and scene arrangement.

For example, a paper airplane takes off between Victorian bookshelves, with books following and forming a tunnel in the air. Then the scene cuts to the microscopic world inside the brain, where neurons glow in translucent nebula-like tissue, like a miniature version of the Milky Way.

Similarly, a paper airplane makes a non-linear flight in a library maze, diving, skidding, and brushing past obstacles. The camera follows it closely, and there is a strong sense of skidding when it turns, without blurring.

The library, which was warm and amber in color just moments before, immediately switched to deep blue and deep purple upon entering the mental world. However, because the camera movement was continuous, the transition became a visual impact rather than abrupt.

Each of these capabilities, taken individually, is a plus. Taken together, they mean that the V6 has begun to possess a complete cinematic language capability.

The last thing that impressed me was not the visuals, but the sound.

After enabling Audio in V6, audio adaptation is significantly enhanced. I tried a scene from a miniature model's perspective: a volcano is erupting in the distance, villagers of the miniature world are scattering and fleeing, and a giant hand holding a transparent bowl falls from the sky, covering the entire village. The transparent bowl vibrates slightly as the air is compressed as it falls.

This is especially true for ASMR scenes. Sounds like rain on a window, the crackling of a campfire, and the turning of pages in a book all contribute to an immersive experience. The audio quality directly determines whether you're watching a video or truly immersed in the scene.

One person producing a film by sheer force is called efficiency; a group of people producing a film by sheer force is called productivity.

If V6 were simply a faster and better AI video model, the story could end here. But PixVerse clearly doesn't intend to stop there.

It is understood that two things were launched simultaneously: Team Plan and Mini Apps.

Team Plan is designed for studios with 2 to 15 people. The core mechanism is simple: the team shares a single points pool.

Instead of each person having their own account and managing their own data, the entire team uses resources uniformly, with permissions assigned according to roles. The boss has a global perspective, the creative director oversees project groups, and editors focus on producing the final product. Materials and templates from individual spaces are synced to the team space with a single click, eliminating the need to transfer files back and forth in the group chat.

It sounds simple, but it solves a very real problem. Previously, our studio used AI video tools where everyone had their own account, and good materials were shared in a group chat. Version control relied on file names and dates. This is essentially no different from transferring PowerPoint presentations via USB drives ten years ago.

What Team Plan does is upgrade AI videos from personal tools to a team production line. One person producing ten videos and choosing the best one is called individual efficiency; five people each producing ten videos and choosing the most outstanding one out of fifty is called team productivity.

Mini Apps take a different approach. They're so simple to use they need no explanation: upload a few product images, and the system automatically edits them into a complete advertising video. No editing skills or knowledge of camera techniques are required.

A Taobao shop owner, a Xiaohongshu blogger, or a street-side milk tea shop owner can all get a product video that can be posted directly within minutes.

The V6 model is a weapon for creators, while Mini Apps are a point-and-shoot camera for everyone. When the barrier to creation is lowered to zero, "burst-out" photography becomes more than just a methodology for creators; it becomes infrastructure that everyone can use.

Speed ​​is the most expensive aspect of image quality.

The image quality is good enough, the speed is fast enough, and the cost is also worth discussing separately.

The standard cost for 720p resolution is approximately $0.04 per second, which is low in the current market, yet it offers top-tier production quality. Memberships purchased before April 7th offer up to a 30% discount, and the points required for production are further reduced by 30%, further lowering the cost of large-scale production.

This reminds me of the shift in photography from film to digital.

In the film era, a roll of film contained 36 exposures. Before pressing the shutter, photographers would spend three minutes mentally composing the shot, because each shot had a cost, but the quality of the shutter itself was the same. How do photographers in the digital age shoot? Because every shot is clear and sufficient, they can take 200 shots in quick succession and then choose the one with the best composition. No one will say this is "unserious," because ultimately, the audience always sees the best one.

V6 ushered in this "digital age" for AI video creation.

PixVerse's understanding of "fast" has long gone beyond the generation speed itself.

From the early high-speed generation, to the real-time interaction of R1, to the strong command understanding of V6 that makes descriptions more natural, and then to Team Plan that brings collaboration efficiency to the team level, this line shows that PixVerse is accelerating in multiple dimensions at the same time: fast generation speed, accurate command understanding, and improved collaboration efficiency.

Supporting all of this is PixVerse's robust model iteration capabilities.

The reason it has consistently maintained its position in the top tier of the rankings is that each generation of its products has been dedicated to tackling the most challenging problems: physical simulation, portrait texture, and cinematic language—each a difficult but correct direction. V6's initial intention was simple: to allow creators to focus their energy on creativity, leaving the rest to the tools.

When a tool makes high-quality content creation readily available in daily life, lowers the barrier to entry to zero, and elevates collaboration efficiency to the team level, it is no longer just a "model." It is becoming the infrastructure of the content era.

PixVerse V6 ushered in an era of explosive AI video production, delivering videos at breakneck speed and with unparalleled quality, ready to be submitted immediately.

This era has only just begun.

Authors: Li Chaofan, Mo Chongyu

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.

In 2006, in Mayang County, Hunan Province, it was raining coldly, and the temperature was only 10 degrees Celsius. A secondhand motorcycle, older than a person, was speeding along the muddy mountain road.

The cyclist was a 19-year-old boy. In order to get a chance to appear on camera, he braved the rain and relentlessly chased after the Hunan TV interview bus for three hours, covering more than 100 kilometers.

When he stopped the reporter, he was covered in mud and shivering from the cold.

Facing the reporter's camera, he pleaded, saying that he knew he was losing face and that even his fellow motorcycle enthusiasts back home were laughing at his naivety, but he really wanted to join a professional motorcycle team.

As long as I can join the fleet, I'm willing to do anything: washing clothes, cooking, or repairing cars.

His name is Zhang Xue, and he is a car repair apprentice who dropped out of junior high school. At that time, he had nothing and couldn't even afford a decent car. He was willing to do anything to touch a real race car.

Twenty years have passed.

Last weekend, the Portimão circuit in Portugal resounded with the "March of the Volunteers." In this WSBK World Superbike Championship, French rider Valentin Debis won both rounds of the middleweight division.

For decades, the podium in this category has been dominated by established European and Japanese automakers such as Ducati and Yamaha.

The motorcycle that knocked them out this time is called the 820RR-RS, equipped with a Chinese-made 819cc three-cylinder engine. And the company that built this motorcycle is called Zhang Xue Motorcycle.

The wild child who shamelessly begged for shelter in the cold rain back then is now the owner of this championship-winning racing team. He owns a motorcycle company with annual revenue of hundreds of millions and has successfully etched his name on world-class racetracks.

It sounds like a domineering CEO wish-fulfillment novel.

Lei Jun, who emerged from a car repair shop

For those who don't ride motorcycles, the name Zhang Xue might be somewhat unfamiliar. If you had to find a point of reference, you could think of him as the Lei Jun of the motorcycle world.

They all possess strong personal brands and massive followings, and their new brands are generating an explosive volume of orders for their first products immediately upon launch. More importantly, they want to prove their capabilities by taking their cars to top-tier racetracks.

However, when Lei Jun entered the car manufacturing industry, he had tens of billions of yuan in cash on hand. Zhang Xue's trump card, on the other hand, was his passion for motorcycles, which had led him from his days working in a repair shop.

Born in 1987 in a rural area of ​​Huaihua, Hunan, Zhang Xue started from a disadvantageous position.

His parents divorced when he was 3 years old, and he grew up with his grandmother. When he was 10, he and his sister lived in a mud house with cracked walls. When it was windy and rainy, the two siblings could only find some old plastic sheets to plug the gaps that leaked water and air.

Later, some media followed him back to his hometown to see that shack and asked him if he felt sad when he lived there.

Zhang Xue said calmly, "When people have no choice, sadness is useless. After experiencing all the hardships, they will feel happy when they encounter even a little bit of sweetness."

This insensitivity to material conditions foreshadowed his later obsessive career in car manufacturing.

The year after being featured on Hunan TV, Zhang Xue took the 60,000 yuan she had saved from repairing cars and went to work for a car team in Siyang, Jiangsu. Reality was often harsh; even after joining the team, there was no guarantee she would get a car to drive. The team wouldn't allocate expensive training resources to someone without connections. Therefore, Zhang Xue's daily work consisted mostly of repairing cars and doing odd jobs.

Fortunately, an experienced driver named Zhang Jixing recognized his tenacity and began to train him privately.

But the longer Zhang Xue stayed in the paddock, the clearer a harsh reality became: motorsports is a sport that requires a lot of money. No matter how good your driving skills are, without top-notch racing cars and substantial financial support, a poor kid will never reach the top podium.

If you can't afford a nice car, then build one yourself.

After shifting her focus from the racetrack to the workshop, Zhang Xue began to delve deeper into the underlying mechanical technology.

When he was at Zhejiang Apollo, he was able to assemble a bunch of scattered parts into an engine that could start, even with his eyes closed, in front of CCTV's cameras. In 2013, he took 20,000 yuan and plunged into Chongqing, the city with the most mature motorcycle supply chain in China. By finding parts in the parts market, assembling them himself, and reselling them, he mastered all aspects of motorcycle manufacturing.

Finally, in 2017, Zhang Xue and several partners founded Kaiyue Motorcycle.

At that time, the domestic mid-to-large displacement motorcycle market was almost monopolized by imported brands, and most domestic brands were still following the old path of reverse engineering and assembly, commonly known as "the horizon." Zhang Xue and Kaiyue took a different path: pursuing the best lightweight and handling among models in the same class.

Although Buick's first mass-produced model, the 500X, had various minor issues, it quickly achieved commercial success in comparison to its competitors. It sold 800 units in its first year, and sales jumped to 3,000 units in its second year, earning the company its first pot of gold.

Refusing to be a profitable assembly plant

Although the cars sold out, Zhang Xue felt uneasy.

Like many domestic brands at the time, the early Buick Excelle used readily available engines sourced from suppliers. Having the core components of the entire vehicle in someone else's hands not only limited the product's performance potential but also left it vulnerable to future supply chain disruptions.

He wants to build a large-displacement engine that truly belongs to China.

However, within the Kaiyue team in 2019, when Zhang Xue proposed developing an 800cc inline twin-cylinder water-cooled engine, the partners expressed clear opposition due to concerns about profit and risk. Research and development is a bottomless pit; once tens of millions are poured in and no results are seen, all the company's hard-earned assets from the previous years will be wiped out.

In order to save the project, Zhang Xue proposed a plan.

As the company owner, he borrowed 10 million yuan from the company to advance research and development. If the project succeeds, the intellectual property rights of the engine will belong to Kaiyue; if it fails, he will personally bear the entire debt.

After a year and a half of tedious research and development, the 800cc engine, codenamed EBS01, successfully ignited in early 2021. Following this, Buick Excelle initiated and successfully developed a 400cc four-cylinder engine capable of reaching 16,000 rpm.

The launch of these two models not only allowed Buick Excelle to shed its "assembly plant" label, but also gave it the confidence to truly compete with leading domestic manufacturers like CFMOTO.

With a mature product in hand, Zhang Xue's innate racing genes once again took over. Leading a team to participate in competitions became an essential way for Zhang Xue to test the reliability of the vehicles.

In 2023, Zhang Xue led the Buick Excelle team to participate in the Dakar Rally, becoming the world's first manufacturer to complete all its cars in its first participation. In the China Taklimakan Rally, he himself was a driver when he was hit by an off-road vehicle, resulting in a brief period of unconsciousness and a broken finger. After regaining consciousness, he still insisted on finishing the day's stage.

However, as the company grew, the logic of capital and Zhang Xue's obsession inevitably clashed.

By early 2024, Buick had grown into a company with annual revenue of hundreds of millions.

The shareholders' demands are very clear: stop those bottomless new engine development projects and focus their efforts on selling existing models, making a quick profit first and then securing their gains.

But in Zhang Xue's view, if a car company stops pushing forward in core technologies, its decline is inevitable. Car manufacturing is a gamble that cannot be stopped, and relying on past achievements to survive goes against his original intention in establishing the company.

Faced with pressure from upper management to step down as general manager, Zhang Xue compromised, stating that he could remain as vice president of R&D. He even wrote a lengthy proposal to persuade the board to continue investing in R&D.

The suggestion was ultimately rejected.

With the budget for building the new engine completely locked up, Zhang Xue felt that staying any longer was pointless.

On March 1, 2024, 37-year-old Zhang Xue posted a resignation letter on her WeChat Moments.

After careful consideration, I have decided to resign and pursue my dreams. In the future, we will be friends and rivals. Goodbye.

The founder relinquished a huge amount of his shares, choosing to leave the star company he founded with nothing.

The only way to fight back is to win the game.

Zhang Xue barely got any rest.

In April 2024, just one month after leaving Kaiyue, he started a new team in Chongqing, and Zhang Xue Motorcycle was officially established.

Having learned from her previous experience of being forced out due to disagreements over funding, Zhang Xue has become wiser this time. The new company secured angel investment from Gaoxin Capital, but Zhang Xue firmly retains 80% of the shares, keeping absolute control over the technology roadmap and racing car planning. He will have complete say in what cars to build and what races to compete in.

Like Xiaomi Auto, this new team is building cars at an extremely fast pace.

In less than a year, the first mass-produced motorcycle, the 500RR, was officially launched. From the first day of its release, Zhang Xue did not treat it as an ordinary commuter motorcycle, but instead focused all her marketing efforts on its racing DNA and extreme handling.

The market responded with overwhelming enthusiasm. In just four months, 10,000 500RR units rolled off the production line. By the end of 2025, the new company's revenue had reached 700 million yuan as expected.

The car's phenomenal sales are not only due to its excellent product quality, but also largely due to Zhang Xue's personal style of doing things.

Nowadays, car company bosses avoid negative news like the plague and try every means to quell the situation, but Zhang Xue insists on doing the opposite. In his live stream, he openly encourages car owners to expose any problems they encounter online.

My only approach is to expose the problem to the light of day and quickly shut it down.

He also used this straightforward, chivalrous spirit to support his users.

Unlike Hong Kong, Macau and Taiwan, in the traffic environment of mainland China, motorcycles are often the least welcome group, and riders are often targeted by cars.

In the second half of 2025, two separate incidents occurred in Jinan and Hefei where Zhang Xue, a motorcycle owner, was maliciously cut off, verbally abused, and even physically assaulted by car drivers. After watching the videos, Zhang Xue publicly called on her company's legal team to proactively contact the motorcycle owner and assist in reporting the incident to the police, ultimately resulting in the two arrogant car drivers being detained.

This protective behavior garnered a large number of loyal fans for Zhang Xue's new brand.

But Zhang Xue's ambition was never just to sell 500RR motorcycles and make a comfortable living. He held a trump card in his hand: a completely self-developed 819cc inline three-cylinder engine.

When Zhang Xue publicly stated that she would use this new platform to compete in top-level international races, she was met with a barrage of ridicule. Someone posted a screenshot online, saying that the biggest joke they'd heard this year was Zhang Xue's 820 trying to compete with the Ducati V4.

Zhang Xue saw the screenshot, posted it online, and added a caption.

I've never lacked ridicule in my life, but I know exactly what I want to do and respect my own heart. We don't have much time to live, so let's enjoy and live life to the fullest within our limited time!

In November 2025, Zhang Xue announced at the Milan Motorcycle Show that Zhang Xue Motorcycle would officially enter the 2026 WSBK World Superbike Championship, competing in the most competitive WorldSSP middleweight class.

Unlike MotoGP, which tests the R&D capabilities of each company's "prototype" bikes regardless of cost, WSBK examines the extreme performance of mass-produced bikes from major automakers.

Zhang Xue set a rather ambitious goal for the team: to stand on the podium in the first year, win a race in the second year, and win the overall championship in the third year.

Just a few months later, on March 28th, at Portimão station in Portugal.

The race car, equipped with a Chinese-made three-cylinder engine and weighing only 168 kilograms, demonstrated cornering stability and straight-line acceleration capabilities comparable to Ducati and Yamaha, under the guidance of veteran French driver Valentin Debis.

In the first race, Debis started from second place and led the entire race, leaving the Ducati V2 and Yamaha R9 behind him, crossing the finish line with a huge advantage of 3.685 seconds.

The same story repeated itself the next day.

Two races over the weekend, two championships in a row. Zhang Xue had originally planned to complete the journey in two years, but finished it in the second leg of her foray into the European racing scene. For the first time, a Chinese motorcycle manufacturer broke the monopoly of European and Japanese powerhouses, snatching the middleweight race championship from them.

On the night of her victory, Zhang Xue, holding the five-star red flag, was moved to tears.

After 20 years of ups and downs, Zhang Xue hasn't changed at all; he's still walking on that muddy mountain path from back then.

It was a cold, rainy day, and the temperature was very low. The other cyclists around him were all laughing at him.

The 19-year-old boy was covered in mud and his lips were purple from the cold. He got up from the ground, picked up the old motorcycle, and hysterically shouted that sentence into the camera.

If you have a dream, go for it! Because of courage, my life is more wonderful!

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.

For the past two years, we've been doing the same thing every day: learning and studying the mystical art of "keyword engineering."

When hiring AI to do work, you always have to provide 800 words of background information, like you're trying to appease an intern with an exceptionally high IQ who experiences intermittent amnesia every day.

This reminds me of the "pre-cast" or "chanting" process that always precedes the unleashing of a special attack in games. In a way, writing prompts, providing context, uploading various files, etc., are like using AI's "pre-cast."

It's not that users have to be perfect every time, but if you can provide these prerequisites, AI will do better.

However, at an offline gathering in Zhongguancun a while ago, APPSO saw an AI office product that was still in testing—it largely abandoned the reliance on "pre-shaking".

The product is called Floatboat.

Shaoqing, co-founder and CEO of Floatboat, walked onto the stage, opened Floatboat, selected a folder containing a CSV table listing the guests attending the event. He then spoke in the AI-generated dialog box next to him: "Generate invitations."

After a while, invitations for each guest appeared.

Up to this point, everything was fine. If you gave the table to ChatGPT, Claude, WorkBuddy, Wukong, or any of today's AIs, they could probably do something similar by writing a single command. But what happened next surprised me.

A new guest has confirmed their attendance, and Shaoqing said, "Update the form."

The CSV file was updated; immediately afterward, a new invitation was automatically generated.

I sat there for two seconds, trying to understand what had just happened:

Floatboat knew there was a relationship between the form and the invitation, and between the actions of "updating the form" and "generating the invitation." Therefore, even though Shaoqing only said the first half of the sentence, Floatboat figured out the second half on its own.

AI is no longer just a tool waiting for instructions; it's becoming increasingly proactive and intelligent, like a helpful assistant who understands your meaning when you say "update."

This moment made me start to take a serious look at this product.

A product that is simple, yet cannot be simply defined.

What is a floatboat? I tried to define it, but I found it difficult.

It has a file manager that looks like macOS Finder, allowing you to browse local files and access iCloud Drive; it supports a wide range of file formats, including Markdown, CSV, Excel, Word, images, and videos, all of which can be previewed and even edited directly.

It has a built-in browser that can open any webpage and even allow an agent to manipulate those webpages;

It has an AI dialogue interface that can connect to Gemini or other models. In this respect, it's somewhat similar to Claude's desktop client, but with more intuitive operation logic than Cowork.

These three things—files, browser, and chat—are arranged side-by-side in a panel format, and can be dragged and combined freely, with a maximum of four columns side-by-side.

If you see a useful image in your browser, you can drag it directly to your local folder to save it; if you have AI generate a report, the report will be written directly to a local file and saved in .md or .docx format, and you can edit these files directly without having to type cmd-c and cmd-v to another location.

Information flows into this environment from all directions, and processed content can also flow out, without being locked in a certain panel.

So what exactly is Floatboat? Is it a file manager? A browser? An AI chat tool? An ambient programming environment?

It is all of these, yet not entirely.

Before Floatboat came along, we were essentially acting as "human APIs" between different software programs, pressing copy and paste hundreds of times a day, opening different software or browser windows, and editing different files.

In the AI-driven era of online work, we've become cyber porters, frantically running between windows.

Floatboat breaks down the barriers between software, allowing all windows to share the same context.

The development team defines the product as a "work environment" rather than an "AI assistant." An assistant only moves when you ask it to, while a work environment is always there. You work in it, and it observes and learns while helping you.

During the communication meeting, someone asked Shaoqing: "Describe your product in one sentence?"

Shaoqing countered: Can you describe ChatGPT in one sentence?

Everyone smiled knowingly. I think he has a point. Some things really can't be summed up in a single sentence, unless you're making a very vertical tool. Floatboat clearly doesn't intend to be vertical.

Over the years as a tech journalist, I've witnessed several generations of such products. The earliest was the era of email plus Office suites, then came various OA systems, and later DingTalk, Lark, and Slack.

Every generation has a product, or a type of product, that carries the same subtext, giving you a strong hint or explicit message: For work, I'm all you need.

In the AI ​​era, Floatboat wants to become that role.

This isn't to exaggerate its importance. On the contrary, no one has ever truly held this position securely throughout history. Lark solved team collaboration, but document manipulation still requires the Office suite. DingTalk perfected the approval process, but the habit of working people using WeChat to discuss work in private has never changed.

The idea of ​​"unifying the market" has been attempted by every generation of products, but no one has ever truly achieved it.

The reason is structural: for this type of product to succeed, the entire organization needs to change. And organizational inertia is the greatest of all inertias. It's useless if you alone think Lark is good; your team, your customers, and your suppliers all have to think it's good for it to work.

Floatboat's strategy is different in one way: it doesn't target organizations, it targets individuals.

The target audience for this product is precisely the most popular concept right now: OPC, which stands for One Person Company.

The leap forward in AI capabilities over the past year has made OPC, a slogan from two years ago, increasingly realistic and feasible. One person, plus three to five agents, can almost replace a small, fledgling business and support team. Whether it's self-media content creators, handling everything from topic selection, writing, layout, and distribution, or e-commerce businesses, handling everything from product selection, listing, customer service, and traffic generation, it's already sufficient.

Floatboat aims to resonate with this group of people. During our experience with APPSO, we tested various scenarios including content creation, data science, and integration with external tools (such as Slackbot). Floatboat met our expectations for content creation, marketing, data analysis, and customer service.

There are currently two design philosophies for AI products. One is "You let go, I'll do it," where the user is put in the back seat, the agent takes full control, and shows the results when it's done. The other is "You do the work, I'll be there," where the user acts as a co-pilot, providing tools and suggestions when appropriate.

Floatboat is closer to the latter, but not entirely. My experience working with Floatboat is that switching between the driver and passenger seats with the AI ​​is smooth and seamless.

After using it for a while, I think Floatboat's approach is feasible. At least at this stage, most people don't trust AI to the point of "just do it, I don't need to watch." If you let a worker hand over an entire project to AI to run on its own, he'll be so anxious he won't be able to sleep…

But if the AI ​​is working on his screen, next to his folders, and he can see the process and correct it at any time, he will feel more at ease.

This is why Floatboat's interface design resembles a traditional computer desktop, displaying the file manager, dialog boxes, and browser/editor all at a glance: familiar elements reduce users' wariness of something new and increase their acceptance.

Working while distilling

Then let's talk about a feature called Combo that Floatboat implemented.

A combo can be a complex skill or a combination of multiple skills. In the context of workflow, it means packaging a workflow into a reusable operation.

Floatboat has a built-in ability to "distill" combos from work results—which is actually very similar to Anthropic's official skill-creator (which is itself a skill).

For example, you might do one thing every week: scrape a few industry reports from the internet, extract summaries, organize them into Markdown documents, and then push them to Notion. After you manually complete this process with the Agent in Floatboat for the first time, a button will appear at the bottom of the dialog box, asking if you want to save this round of operations as a combo.

Alternatively, you could proactively tell Floatboat, "Let's organize our current work methods, thinking, and logic into a skill."

The next time you encounter a similar task, Floatboat will automatically recommend this combo to you, allowing you to start it with a single click.

The most interesting thing about this is that you don't need to "design" a workflow in advance; you just need to do your work normally. As you work, Floatboat will automatically "distill" your work habits and operating methods, and distill them into a set of guiding principles.

Shaoqing told APPSO that the design of the Combo capability is to fulfill the core expectation of most users for agent products today: self-evolution.

"When the agent can perceive 80% of your operations, it has the ability to self-evolve." Combo's automatic accumulation mechanism is the first step in doing this.

The era of peddling "cheat codes" is coming to an end. You no longer need to memorize tedious spells like a magician, saving cheat codes in a dedicated folder or in the background of an AI tool. Through Combo, Floatboat allows users to distill their most frequently performed daily actions into their own unique "crafts" and digital assets.

Of course, Floatboat also has a Combo marketplace where you can upload your own useful combos, and download those created by others. The official website also provides some ready-made ones.

However, this Combo system still has shortcomings.

Any office software that claims to dominate the market, or an AI system that claims to "understand you better the more you use it," still faces the hurdle of a cold start: just like how Google Docs' initial resume templates, while comprehensive and excellent, still require each job seeker to adjust and modify them to suit their own needs.

Combo's automatic learning mechanism makes sense in theory: the more you use it, the better it learns, and the more its recommended workflows will suit you. But there's a prerequisite: you need to invest time teaching it from scratch, and most people don't have that patience; they want it to be usable right out of the box.

As a media editor, my daily work involves reading a large amount of material, communicating with authors about topics, revising manuscripts, and occasionally writing long articles myself. These tasks are very granular and fragmented, and do not match the official templates (which are more focused on standardized report generation and data processing).

In my specific use, I saved several different content production paths as different combos: one is to react quickly to external news, another is to write based on interview Q&A outlines, and yet another is to conduct research on complex topics, compile materials, and then write original content.

Of course, this isn't a problem with the Combo itself. For the vast majority of people, whether their work involves document writing, report processing, PPT creation, data organization, administrative work, or even the more complex "one-person developer + marketer + customer service" model, it's perfectly usable whether they create their own Combo or make minor tweaks to Floatboat's official Combo.

AI tools are not a panacea for all work—no matter how wonderful a tool promotes itself, today's users should be aware of this. As mentioned earlier, Floatboat is a "work environment"; its capabilities are sufficient to enhance human performance, but its effectiveness still depends on human intervention.

Next, let's talk about the differences between Floatboat and other "Cowork-like" products: the most noticeable difference is that Floatboat's workflow is very fast. Taking file operations and content generation as examples, Floatboat, driven by the Gemini 3.1 Pro model, takes about one-third the time I usually use with Cowork/Claude Code CLI to perform file operations (batch renaming/formatting, filling markdown, etc.).

Gemini is a veteran at "pleasing users," so recently Floatboat has added support for two of Claude's latest models, Sonnet and Opus 4.6.

Gemini is sufficient for most of Floatboat's main office scenarios (copy generation, spreadsheet processing, information organization), and the writing results are also quite good. If it doesn't suit your preferences, switching to the Claude model is also fine. If you notice that Floatboat's attempt to cater to users is too strong, you can occasionally emphasize this during your work process. Don't blindly follow the lead; instead, critically consider the generated results and even the user's input.

Additionally, you can take full advantage of the Combo generation capabilities and incorporate these techniques into the core guiding principles of Floatboat.

Another small design detail worth mentioning is that Floatboat can be integrated into Lark and Telegram. You can send it a message directly in the chat tool without opening its client, and it will perform the task for you in the background—this feature is called Claw mode, which should be self-explanatory.

Beyond the product itself, the Floatboat team is working on something much further ahead.

They open-sourced a license called Selfware, whose core concept can be summarized in one sentence: A file is an app.

What does this mean? Imagine you painstakingly create a research report using AI and send it to a colleague, who receives a Word document or a .md file. The file contains the final results, but the most crucial experience you gained—such as which resources you used, the logic the AI ​​ran, how many times you made changes, and why—is not saved.

Selfware aims to solve this problem. A .self file contains not only data but also logic and structure. Your colleague can open it, continue editing, and guide the agent to follow your intended path. The file comes with its own working environment.

This idea aligns with the current enthusiasm in the AI ​​development community for files like CLAUDE/SKILL.md and cursor rules. Everyone is discovering that text files can be used to "program" AI behavior; a single .md file can define an agent's personality, working methods, and output style.

But Selfware went a step further: those .md files are instructions, telling the agent what to do; Selfware wanted to be an execution unit, where the file itself could run, and it wouldn't depend on any specific platform.

This is actually somewhat similar to Jupyter Notebook, packaging code, data, and output together; it's also similar to Docker, making the runtime environment a distributable unit—Selfware replaces the scenario with agent collaboration. It's not a concept invented from scratch, but its reintroduction in the agent era truly addresses a real pain point.

However, ultimately, the success of any protocol depends on adoption rates. Currently, Selfware primarily operates within Floatboat's own ecosystem. "A file is an app" is an interesting concept, but there's a long way to go from concept to widely adopted standard.

Another noteworthy point is IACT (Inline Action-Clicked Text), another open-source license from Floatboat. It does something smaller but very practical: based on Markdown syntax, it directly adds clickable inline links/buttons to the AI ​​dialogue generation results. The "actionable content" in the generated results will automatically have this button appended, which the user can simply click.

This seemingly minor interaction improvement does reduce friction in use. Claude was probably the first to create a similar experience, but many of Claude's "good things" are closed-source. Floatboat open-sourced IACT, allowing other products to fully utilize it.

Now some similar products, such as WorkBuddy, are also doing something similar, but as far as I know, Floatboat was the first to propose this concept and formalize it.

When working, being happy is the most important thing.

The name Floatboat comes from the English idiom "whatever floats your boat," which roughly means "whatever makes you happy."

Shaoqing said that they hope their products will give people a feeling of floating in the AI ​​era, so that they are not swept along.

That's a good vision. But can Floatboat become the "I'm all you need" product of this era? To be honest, APPSO still can't give a clear answer.

After all, as everyone has seen, each generation of office products that attempted to do this eventually became just one of the tools in the toolbox, rather than the only one.

But it's too early to make a judgment today.

A product doesn't need to unify everyone's work style to be considered successful. If it can free up an hour each day for a group of people—those "OPC" workers who do the work of five people and act as intermediaries between software programs—to do things that truly require brainpower, then it is already worth existing.

For most ordinary people, it is indeed quite tiring for one person to do all the work of a company.

But what's exciting about Floatboat is that it gives an individual the confidence and assurance to run a company.

Not everyone can become an OPC; you first need a good "PC." Floatboat is betting that it will become that PC.

#Welcome to follow iFanr's official WeChat account: iFanr (WeChat ID: ifanr), where more exciting content will be presented to you as soon as possible.