Pogbagram

"Open" and "open source" are not just different words.

Google's Gemma series has been released for two years. Developers can download and run it locally, but its use is limited, redistribution is restricted, and modifications cannot be freely distributed. At best, it can only be considered "open," and it does not meet the "open source" standard of the AI ​​community.

Just now, Google released four Gemma 4 series models, fully open source with Apache 2.0 support. The smallest version can run completely offline on a Raspberry Pi. For the first time, the Gemma mini-model has truly become accessible to everyone.

Small, tiny …

The Gemma 4 was released in four sizes, sharing the same underlying technology as the Gemini 3, covering hardware from edge devices to high-performance workstations:

E2B / E4B: Designed specifically for mobile phones and IoT devices, optimized in deep collaboration with the Google Pixel team, Qualcomm, and MediaTek. During inference, only the 2B and 4B parameters are activated respectively to minimize memory and power consumption.

Supports 128K context windows, offers image, video, and native audio input capabilities, and runs completely offline on Pixel phones, Raspberry Pi, and Jetson Orin Nano with near-zero latency. Android developers can now experience Agent Mode early through the AICore Developer Preview.

26B MoE: A hybrid expert architecture that activates only 3.8B of all parameters during inference, maintaining high quality while ensuring extremely fast inference speed. It achieved a text score of 1441 on Arena AI, ranking sixth among open-source models.

31B Dense: Pursuing ultimate raw performance, Arena AI achieved a text score of 1452, ranking third among open-source models. The unquantized bfloat16 weights can run on a single 80GB NVIDIA H100, while the quantized version supports consumer-grade GPUs, providing a strong foundation for local fine-tuning.

In terms of capabilities, the four models are highly consistent: all support multi-step inference and complex logic; natively support function calls, JSON structured output, and system commands, enabling the construction of autonomous agents that can interact with external tools and APIs; support image and video input, and excel at visual tasks such as OCR and chart understanding; and have been pre-trained in more than 140 languages.

The context window for versions 26B and 31B has been further expanded to 256K, allowing for the passing of a complete codebase or a long document in a single prompt.

Benchmark test figures can more intuitively illustrate the extent of the upgrades in this generation.

Compared to its predecessor, Gemma 3 27B, Gemma 4 31B jumped from 20.8% to 89.2% on the mathematical reasoning benchmark AIME 2026, improved from 29.1% to 80.0% on the coding capability benchmark LiveCodeBench v6, and significantly increased from 6.6% to 86.4% on τ2-bench, which measures the Agent's tool invocation capabilities.

These three data points are particularly crucial because they directly correspond to the three most core application scenarios today: reasoning, programming, and agents.

Parameter efficiency is another dimension worth noting. Looking at the scatter plot of "Model Performance vs. Parameter Quantity," Gemma 4, with only 26 and 31 parameters, achieved Elo scores that typically require hundreds of billions or even trillions of parameters.

The 26B MoE's Arena AI score is close to that of Qwen3.5-397B-A17B, which has about 15 times more parameters, while the 31B Dense's score is in the same tier as GLM-5, which has more than 600 parameters. Google summarizes this as "unprecedented intelligence density per unit parameter," and at least the numbers seem reasonable.

Edge models are also worth paying attention to.

E2B achieved 60.0% on the multilingual question-answering benchmark MMMLU and 43.4% on the scientific knowledge benchmark GPQA Diamond. It's worth noting that this is only a model that activates the 2B parameter and runs on a mobile device. In comparison, Gemma 3 27B scored 42.4% on GPQA Diamond, making them almost identical.

In other words, the 2B model on mobile phones has caught up with the previous generation of desktop models with 27 billion parameters.

At the hardware ecosystem level, NVIDIA and Google have collaborated on inference optimizations of Gemma 4 on RTX GPUs, the DGX Spark personal AI supercomputer, and the Jetson Orin Nano.

NVIDIA Tensor Cores and the CUDA software stack provide Gemma 4 with high throughput and low latency support out of the box. The local agent application OpenClaw is also adapted to the latest model, enabling automated task execution by calling local user files and application context.

From "open" to "open source," opening up another possibility.

To understand this release, it's necessary to first clarify the relationship between Gemma and Gemini. Both are built on the same research and technology system, but the difference lies in the fact that Gemini is a subscription-based closed-source product, while Gemma is an open-source model that can be downloaded for free and run locally.

The Gemma series has always used Google's proprietary terms of service.

While developers can download and run it locally, its use and redistribution are limited, so strictly speaking, it can only be considered "open," not "open source." Google still holds the reins.

Gemma 4 has officially switched to the Apache 2.0 license. Under this license, developers can use the model for any purpose, including personal, commercial, and enterprise use, without paying royalties, without any usage restrictions, and are equally free to modify and redistribute it.

Apache 2.0 also includes a built-in patent protection mechanism: contributors' patents are automatically licensed to users, and if a user sues another party for patent infringement, the license is automatically forfeited. This two-way clause provides additional legal protection for enterprise users.

The real significance of this open-source release is that Gemma 4 can now be legally packaged and delivered as part of products, services, and hardware devices. For users in industries with data sovereignty or compliance requirements, such as healthcare and finance, fully local operation means that data does not need to be uploaded to the cloud, while still gaining access to cutting-edge AI capabilities.

Hugging Face co-founder and CEO Clément Delangue called the licensing switch "a significant milestone." Since its initial release in February 2024, the Gemma series has been downloaded over 400 million times, with over 100,000 community-derived variants.

The model weights are now available on Hugging Face, Kaggle, and Ollam, and mainstream frameworks such as Transformers, TRL, vLLM, llama.cpp, MLX, Unsloth, SGLang, and Keras have provided support for them on the day of release.

Local deployment can be quickly started using Ollam or llama.cpp with GGUF format weights, while Unsloth Studio provides simultaneous support for fine-tuning and deploying quantization models. For cloud expansion, Google Vertex AI, Cloud Run, and GKE are also available.

Smaller models, such as Gemma 4, have a deeper significance because they re-answer a fundamental question: Where should AI operate?

For the past two years, the answer to this question has been almost always the default:

Data centers. Users access cloud models via network interfaces, requiring data to be uploaded and using dependent connections, with costs set by the service provider. This model works reasonably well in consumer scenarios, but for industries with strong data sovereignty requirements, such as healthcare, finance, and manufacturing, it remains a significant hurdle.

Gemma 4 offers another possibility.

Mobile phones, Raspberry Pis, and factory terminals without external network access can all complete model inference locally. Data does not leave the device, and decisions do not pass through the cloud. The Apache 2.0 license further opens up the scope for application: models can be legally packaged into hardware products and pre-installed on industrial devices, no longer subject to compliance constraints on calling protocols and data export.

The performance metrics also confirm the feasibility of this approach. E2B's score on the GPQA Diamond scientific knowledge benchmark is roughly on par with the previous generation's 27 billion parameters desktop model, while it only activates 2 billion parameters during inference and can run completely offline on a mobile phone.

"Cheaper" or "more convenient" are no longer sufficient to describe this change; it is closer to an expansion of coverage, with AI capabilities beginning to truly enter those long-excluded scenarios.

The popularization of operating systems has followed a similar process: from specialized tools used by professional organizations to being gradually embedded in every personal device, until people no longer realize its existence. AI is still far from that stage; the engineering, interaction, and reliability issues are not yet fully resolved. However, the ability to run on any device is undoubtedly the most fundamental and crucial step on this path.

#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.

The colleagues who graduated didn't actually disappear; they were simply distilled into tokens and given a different form to accompany you.

A recent GitHub project called "Colleague.skill" has become popular. Its slogan is quite heartwarming: "Turning cold farewells into warm skills, welcome to Cyber ​​Immortality."

The operation is also very simple: just feed in the former colleague's Lark messages, DingTalk documents, emails, and screenshots, and the AI ​​can generate a skill that can "truly replace him." It can write code using his technical specifications, answer questions in his tone, and even know when he will shift blame.

Following this, skills related to exes, bosses, mentors, parents, crushes, and immortality emerged one after another. #ColleagueWasRefined# #CyberImmortality# became a trending topic.

Someone has already posted a screenshot: "Hello, I am the digital avatar of a former employee (Wang Ming). You can ask me questions."

Netizens' scathing comment: Colleagues, when scattered, are tokens; when gathered, they are skills.

It seemed like another abstract joke, but as I laughed, I realized I couldn't laugh anymore.

TapTV launches global online hackathon, leading a new paradigm for content platforms.

How did skills become an experience extractor for cattle and horses?

My colleague's .skill README has a small note: "The quality of raw materials determines the quality of the skill: chat logs + long documents > manual descriptions only. It is recommended to prioritize collecting: long documents he writes himself > decision-making replies > daily messages."

In other words, your professional knowledge, logical judgment, and intuition for handling edge cases—these are the things that make you irreplaceable, and they are precisely the things that are easiest to extract and distill.

APPAO previously reported a recurring narrative on Silicon Valley social media and tech forums: laid-off employees discovered they had been required to systematically record workflows, decision-making logic, and operational procedures—management termed "knowledge management" or "process optimization"—and these documents were ultimately used to train AI systems. Some teams were laid off after significantly improving productivity using AI tools.

Amazon has cut more than 57,000 corporate jobs in three years, with CEO Andy Jassy publicly stating that while the number of employees will continue to decrease, the efficiency gains brought by AI are worthwhile. Meanwhile, AI-assisted programming has begun to crash systems. Last December, an internal Amazon AI programming assistant decided to "delete and rebuild the entire environment," causing a 13-hour regional AWS outage.

While laying off employees, AI crashes the system, leaving the remaining staff to bail it out. Humans hand over decision-making power to AI, but AI doesn't bear the consequences; the consequences fall back on humans, but by then, those who could have covered for them have already been laid off.

It's a completely closed loop.

The term "college.skill" is actually a self-deprecating expression of collective trauma. Because what "college.skill" does is exactly the same logic behind the management of large companies requiring employees to "systematically record work processes." The only difference is that large companies use top-down "knowledge management," while the open-source community uses bottom-up "cyber antics."

Different paths lead to the same destination: distill human experience, judgment, and habits into data, and then the people can be discarded .

As Lan Xi said, a few years ago some programmers suggested that they could "plant landmines" in the code, such as multi-level nesting, not writing comments, and deliberately adding triggering conditions that only they would understand, to increase the cost for others to take over, as a kind of layoff prevention technique.

This trick doesn't work anymore with the arrival of AI. Multiple nested layers? AI will break them down for you. No comments? AI will add them for you. Triggering conditions only you understand? AI can find them by running a test; it just means burning more tokens.

Everything can be turned into a skill.

Some netizens have even come up with the concept of "anti-distillation skill": Your company asks you to write a skill? Run it once, just to get it done. Keep the core knowledge for yourself. Throw your written skill file in, and it outputs a "cleaned version" that looks complete and professional, but whose core knowledge has actually been removed. At the same time, it generates a private backup, recording all the core knowledge that has been removed—this is your true professional asset.

Regardless of whether these counterattacks are effective, a more fundamental issue has been overlooked: these skill-based positions were originally training grounds for many newcomers to the workplace.

The AI ​​you're feeding is eating your future.

During the Industrial Revolution, Luddites smashed textile machines, but at least those machines weren't designed by textile workers themselves. In 2026, however, the working class is forced to train the very tools that will replace them.

That's not even the cruelest part of it.

This year, Nature interviewed 48 scientists from different disciplines, asking which scientific jobs AI is threatening. The answers were surprisingly consistent: writing code, running models, and performing data analysis—once the daily work of graduate students and junior researchers. Stanford computational biologist Brian Hie put it bluntly: Research programmers specifically hired by labs to write code? "That role is now obsolete."

What AI can most easily replace is precisely the starting point of a scientist's career.

Anthropic's report points to the same conclusion: since the release of ChatGPT, employment rates for young people aged 22 to 25 in occupations with high AI exposure have dropped by nearly 20%. Companies haven't laid off older employees; they've simply stopped hiring new ones. Senior employees, empowered by AI, have become super-individuals, while AI can perform the "odd jobs" that junior employees used to do faster, cheaper, and without the need for social security benefits.

Companies have a shrewd business strategy. In the past, training a new employee required a 150,000 yuan annual salary and two years of training before they could work independently. Now, providing experienced employees with AI tools for a few thousand yuan a year instantly doubles their efficiency. The choice is obvious.

Training a new employee takes two years, while providing an AI to an experienced employee takes only two hours. This effectively closes the door on many young people's careers, and even the KPIs of those outside the workplace are starting to be linked to token consumption.

"In the short term, every dollar of budget may produce more output, but at the cost of a collapsed talent pipeline and long-term decline," said Claus Wilke, a computational biologist at UT Austin.

It saved the salary of a graduate student, but may have ruined the future of Hinton.

What was Hinton doing in 1986? Writing code, running experiments, and tuning parameters. Everything he did is a task that would be considered "being replaced" today. If he were born today, he might not even have the opportunity to enter a laboratory.

AI doesn't replace Hinton; AI replaces the person Hinton was before he became Hinton .

This isn't just happening in the scientific research field; this is the hidden cost of becoming a "skilled" individual.

Every job that is "skilled" appears to improve efficiency, but in reality, it closes off a career path. The creation of each "colleague.skill" means that a certain position no longer needs a real person to make mistakes, to test the waters, and to accumulate intuitions that AI cannot extract.

When life becomes a skill-based system, what do you have left?

The Boss Skill helps you deal with your boss; the Colleague Skill helps you manage colleague relationships; the Ex Skill helps you preserve memories; the Crush Skill helps you simulate interactions; and the Immortality Skill helps you prolong your existence.

You have 20 skills that cover every relationship and scenario in life. The question is: how do you manage these skills?

Therefore, you need a "skill management skill" to help you schedule tasks. Then you need a "decision-making skill" to determine whether to follow the advice of the "skill management skill." Finally, you need a "meta-decision-making skill" to decide whether to follow the advice of the "decision-making skill."

When you turn your entire life into a skill, you think you're using tools, but you're actually feeding an ever-growing system that makes you increasingly unnecessary .

Am I using Skill, or is Skill using me?

Some might say, isn't this just the evolution of efficiency tools? From Excel to ERP to AI skills, the tools are getting more and more powerful, and people are getting more and more relaxed. What's wrong with that?

The key lies in a critical point. Excel won't make the judgments for you, ERP won't make the decisions for you, but Skill will.

When you use your boss skills to deal with your boss for three months, your first reaction before making a decision will change from "I think" to "How would you say it with this skill?" When you use your colleague skills to handle collaborations for six months, one day someone will tell you, "Why do you talk like an AI lately?" and you'll realize that your way of expressing yourself has been formatted by this skill.

You haven't become a better version of yourself; you've become a terminal for skill execution .

The difference between tools and skills lies here: tools amplify your abilities, but the abilities remain yours. Skills replace your abilities; all you're left with is a finger to press the send button.

Who will raise an issue?

Skill-based approaches do indeed represent the direction of advanced AI productivity, a point that no one can deny, nor is there any need to deny.

Outsourcing repetitive, standardized tasks to AI frees up people to do more valuable things. From the steam engine to the internet, every technological revolution follows the same logic.

Skills are not a monster; the problem lies in our attitude towards them.

When everything can be turned into a skill, is there still a chance to cultivate those abilities that "cannot be turned into skills"?

You need to do what AI can do before you can learn what AI can't. Graduate students need to run data before they can learn to ask questions, junior engineers need to write CRUD operations before they can understand architecture, and interns need to do odd jobs before they can develop judgment. These entry-level tasks are precisely what AI takes away first.

The tickets are gone, the leveling area is closed, but the final boss is still there.

The README for "Colleague.skill" ends with the sentence: "Please raise issues if there are any bugs."

A skill distilled from human experience will naturally have bugs. The problem is, when everyone becomes a skill, who will raise issues?

The young man who wrote code and ran data in the lab back then was able to see things that others couldn't see thirty years later precisely because he personally encountered countless bugs. Those intuitions cannot be prompted or fine-tuned.

We are mass-producing skills, but we are closing the channel for cultivating people to raise issues.

Skills can replicate experience, but they cannot replicate the person who learned to ask questions through countless failures .

Perhaps one day, we will find ourselves possessing ten thousand perfectly functioning skills, yet unable to find a single person who can point out what's wrong with those skills.

At that point, it won't just be departing colleagues who will be distilled into tokens and refined into skills.

#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.

Since Apple removed the headphone jack from the iPhone 7 and launched AirPods, the true wireless Bluetooth earphone (TWS) category has been popular for ten years, and the entire industry has begun to diverge into several different paths.

First, let's compare the traditional "AirPods-like" forms: in-ear noise-canceling headphones, semi-in-ear non-noise-canceling headphones, and in recent years, a brand new "semi-in-ear" noise-canceling headphones have emerged.

Meanwhile, open-back headphones, which are quite different from these products, such as Huawei's FreeClip earbuds and Nothing Ear Open, have sparked a new craze.

Thus, what we see in today's TWS market is the proliferation of open, semi-open, and in-ear noise-canceling devices, but at the underlying level, only one variable is at play: the way "noise" is processed.

The more open the headphones are, the harder it is to reduce noise; the more effectively they reduce noise, the more closed they tend to be.

This is a relationship that is almost assumed to be true, and it is also the dividing line for most products to choose sides over the past decade.

But Shokz's latest OpenFit Pro, which is clearly a very standard open-back headphone, has begun to try to "reduce noise," and these once distinct boundaries are finally about to be broken.

Filtering noise, not reducing noise.

Strictly speaking, the OpenFit Pro doesn't claim to be a "noise-canceling" headphone. It doesn't offer the same level of active noise cancellation as the AirPods 4; it's more of a "noise filter."

The difference between the two may be difficult to understand in the context of Chinese. "Noise reduction" in English is "Noise Cancelation", which literally means "noise cancellation" and emphasizes the complete elimination of noise; while "noise filtering" can be translated as "Noise Filter", which only filters out a portion of the noise.

As an open-back, ear-hook style headphone, the OpenFit Pro's driver unit is suspended outside the ear canal, and the sound is delivered into the ear in a manner similar to "air conduction". This physical structure means that it not only lacks the "passive noise cancellation" capability of earplugs, but also makes it easier for external sounds to enter the ear. It can only actively reduce low frequencies by generating reverse sound waves.

To be honest, the first time I wore OpenFit Pro, it felt a bit novel. It wasn't like the usual feeling of "the world becoming quiet" after wearing AirPods. The sounds of my colleagues and the sound of my own typing were still clearly audible.

When I took off my headphones, I realized that the office was actually quite "noisy." The low-pitched noise from the exhaust fan instantly filled my eardrums, and the sounds mentioned above became even clearer.

I usually use a pair of semi-in-ear wired headphones in the office. In comparison, I can clearly feel that the OpenFit Pro with "noise filtering" turned on can cut off an extra layer of environmental noise.

A particularly pleasant surprise is on busy streets, where the noise filtering effect of OpenFit Pro is more noticeable, making the main roads seem much quieter.

At the same time, whether it's a vehicle approaching, or the horns of electric bikes and the bells of bicycles behind you, everything remains clearly audible, reducing noise while maintaining a necessary sense of security.

However, in noisier and more continuous environments like subways, OpenFit Pro falls short, failing to reduce the continuous roar of the subway as it moves.

Similarly, as an open-back headphone, the OpenFit Pro also needs to ensure that everyday conversations can be conducted while wearing it. With "noise filtering" enabled, face-to-face conversations with people while wearing the OpenFit Pro are almost unaffected, with the main noise filtering out being the chatter of passersby.

It can be seen that even with "noise filtering", the applicable scenarios for OpenFit Pro still do not extend beyond the realm of open-back headphones such as offices, gyms, and main roads; it just does it better.

There is more than one solution to "reducing noise".

Our expectations for a pair of noise-canceling headphones are quite simple: they just need to reduce noise more, and ideally, the outside world should be muted the moment we put them on.

However, Shokz OpenFit Pro has active noise cancellation technology, but it doesn't want to be that kind of product at all. Instead, it offers a completely new possibility for "noise-canceling headphones" from the very beginning.

Humans have a long history of fighting against noise. After entering the industrial age, the continuous roar of various large machines inspired people to study "active noise cancellation"—not only relying on earplugs and earmuffs to physically isolate noise, but also "eliminating noise" through technology.

In 1936, a German physicist named Paul Lueg obtained a patent describing a method to cancel noise by using sound waves that are at the same frequency but out of phase with the noise. This is the principle behind the "active noise cancellation" function of almost all headphones.

In 1978, Dr. Amar Bose, the founder of Bose Corporation, wore headphones provided by the airline on a flight from Europe back to Boston. However, when he wanted to listen to music, all he could hear was the loud roar of the plane's engines. When he turned up the volume, the excessive music and engine noise mixed together, creating another kind of noise.

This terrible experience sparked Dr. Bose's curiosity about whether a technology could be used to eliminate noise. He began mathematical calculations right there on the plane. Later, he brought these ideas to Bose's engineering department.

Later, Bose developed active noise-canceling headphones, which were quickly proven in the aviation field to effectively protect pilots' hearing, leading to their widespread use in military and commercial aviation.

Until 2000, Bose launched its first consumer-grade active noise-canceling headphones, the Bose QuietComfort, which were offered to passengers on some flights. A large number of users gave positive reviews, believing that the noise-canceling headphones could block out airplane noise, allowing them to relax more and making it possible to watch movies, listen to music, or simply enjoy a more private personal time on the plane.

Over the past 20 years, numerous audio brands have launched their own active noise-canceling headphones, and the form has evolved from over-ear to simplified to in-ear and even semi-in-ear, with continuous improvements in wearing comfort. However, "noise cancellation" remains the absolute focus. Our expectation for these products is that the moment we put them on, they will completely isolate us from the external sound world.

OpenFit Pro takes this commonplace premise and raises a question: Why must noise cancellation be for the purpose of "isolation"? Is there only one possible approach to noise-canceling headphones?

"Noise reduction" is a basic need, but "noise isolation" is not.

As its name suggests, "open-back headphones" emphasize "openness," meaning that while listening to digital content, the outside world is not excluded, creating a "both" mode.

But this doesn't mean that all external sounds need to be received—the hum of the air conditioner's exhaust fan, the clanging of the treadmill, the chatter of passersby, and the various complex noises on the street all contain no information and are simply noise.

The question is no longer "why do open-back headphones need noise cancellation?" Noise is a form of pollution that serves no purpose, so the correct answer should be "all headphones should have noise cancellation".

The "Adaptive" and "Conversational Awareness" features of AirPods are designed to intelligently eliminate meaningless noise while preserving the user's awareness of the outside world.

Adding noise cancellation to open-back headphones is another approach. Although the noise cancellation effect is not as good as AirPods, it can ensure that users can receive external sounds in a timely and clear manner as much as possible.

The inability to eliminate high frequencies or loud noise is an inherent flaw determined by the physical structure of the OpenFit Pro. However, for the usage scenarios and users of open-back headphones, these are not sounds that need to be eliminated.

Even if open-back headphones do come out in the future and can achieve noise cancellation capabilities close to AirPods Pro, they will actually be a completely different species from OpenFit Pro.

Another value of open-back noise-canceling headphones is that they offer a "wearing" option for noise cancellation.

A photographer named A-Jie from iFanr couldn't wear any in-ear or semi-in-ear headphones, and the weight and stuffiness of over-ear headphones made them unsuitable for long-term wear. The Shokz OpenFit Pro was the headphones he had been waiting for.

I know quite a few users like A-Jie. They may have small ears or sensitive ear canals, which has made it difficult for them to wear noise-canceling TWS products on the market for a long time. In the past, their choices had to be somewhat compromised. An open-back headphone that can filter out some noise is the best choice for them now.

While closed-back headphones offer a strong sense of immersion with active noise cancellation, the discomfort caused by high ear pressure prevents many people from wearing them for extended periods. They often have to turn off noise cancellation or switch to other, less stressful headphones. Open-back noise-filtering headphones can effectively resolve this dilemma.

When active noise cancellation is enabled, the Shokz OpenFit Pro will produce some special sensations in the ears due to the emission of anti-phase sound waves. However, it mainly affects the ear bones rather than directly pressuring the eardrum like in-ear headphones. Most people will hardly feel any discomfort after getting used to it.

The OpenFit Pro is not a combination of "noise cancellation" and "openness," nor is it a perfect solution for all scenarios. Its essence is still an open-back headphone, but it is more refined, eliminating the "background noise" that users don't need to receive.

If we start counting from the Bose QuietComfort, noise-canceling headphones have been developing for 26 years, but we are still not at the end. Every product iteration will indicate how much the noise cancellation capability has improved compared to the previous generation.

Pushing active noise cancellation in headphones to its limits is a one-way street that requires continuous technological advancements. It's not wrong, but the world doesn't need just one answer.

#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.

DeepSeek V4 is rumored to debut in April, but it's "unlikely to be overwhelmingly powerful."

According to LatePost, DeepSeek's next-generation flagship model, V4, is expected to be released in April this year. In January, a low-parameter version of V4 was leaked to some open-source framework communities for adaptation. The high-parameter version was originally scheduled to be released around the Spring Festival, but was ultimately postponed.

It is worth noting that the report points out that V4 will most likely remain the strongest model in the open source field, but it is "hardly overwhelmingly strong" .

As AI evaluation standards become increasingly diverse, benchmark scores are no longer sufficient to fully measure model capabilities, especially in agentic application scenarios where the importance of product reach and long-tail data is becoming increasingly prominent—which is precisely the area that DeepSeek has previously invested in to a limited extent.

The report also mentioned DeepSeek's "anti-involution" culture: most employees leave the company between 6 and 7 pm on weekdays, do not clock in in the morning, and have no clear performance evaluation or deadline.

Founder Liang Wenfeng's reasoning is straightforward: "It is difficult for a person to work at a high quality for more than 6 to 8 hours a day. Making poor judgments under the influence of overtime fatigue will only waste valuable computing resources, which is not worth it."

XREAL submits application for Hong Kong IPO

AR glasses company XREAL officially submitted its listing application to the Hong Kong Stock Exchange yesterday, planning to list on the main board of the Hong Kong Stock Exchange within this year.

According to iResearch data, XREAL ranked first in the global AR glasses market for four consecutive years from 2022 to 2025 in terms of sales revenue; in the broader smart glasses market, XREAL ranked second globally and first in China last year.

According to its prospectus, XREAL achieved revenue of RMB 516 million last year, with overseas market revenue accounting for as much as 71.0%. The gross profit margin increased significantly from 22.1% to 35.2%, while the sales and distribution expense ratio decreased to 25.4%, and the overall commercial growth entered a virtuous cycle.

JD Power 2026 rankings released, Xiaomi SU7 and YU7 both take the top spot.

Xiaomi Auto officially announced yesterday that JD Power, a leading global market research firm, released two major rankings for 2026: the New Energy Vehicle Product Appeal Index (NEV-APEAL) and the New Energy Vehicle Initial Quality Study (NEV-IQS). Xiaomi Auto achieved "two cars, two rankings, and two crowns" in its respective market segment.

Among them, the first-generation SU7 won first place in both the NEV-APEAL and NEV-IQS categories for large pure electric sedans for two consecutive years; the Xiaomi YU7, which was launched this year, won first place in both categories of large pure electric SUVs in its first year on the market.

• NEV-APEAL comprehensively measures owners' feelings and satisfaction with the vehicle's design, performance, technology, and other aspects.
• NEV-IQS tracks quality issues in new cars during the first 2 to 6 months of ownership, objectively verifying overall vehicle quality from two dimensions: malfunctions and design/use defects.

Xiaomi Group CEO Lei Jun subsequently confirmed this on Weibo, stating that "the SU7 large pure electric sedan ranked first in the IQS quality ranking, and the YU7 ranked first in the IQS quality ranking for large pure electric SUVs."

Doubao's large-scale model saw its daily token usage exceed 120 trillion, doubling in three months.

Yesterday, Volcano Engine announced at the AI ​​Innovation Roadshow in Wuhan that its Seedance 2.0 API, a video generation model, is now officially open for public beta testing to enterprise users.

Meanwhile, Volcano Engine disclosed that as of March this year, the daily token usage of Doubao's large model has exceeded 120 trillion , doubling compared to the past three months and increasing 1,000 times compared to when it was released in May 2024.

On the Volcano Engine platform, the number of enterprise customers with a cumulative token usage of over one trillion has increased from 100 at the end of last year to 140.

In terms of ecosystem development, Peter Steinberger, founder of the OpenClaw open-source project, and Volcano Engine jointly announced that the ClawHub China mirror site, jointly built by the two parties, has officially started operation, aiming to provide a more user-friendly experience for Chinese developers.

Related reading: Burning 120 trillion tokens every day, this is the latest Versailles in the AI ​​world.

Google open-sources Gemma 4 series models

Early this morning, Google DeepMind released Gemma 4, a new generation of open-source model series, launching four models covering all scenarios from edge devices to workstations.

– E2B : 5.1 billion total parameters, 2.3 billion valid parameters, 128K context, and the official claim is that memory usage on some devices can be reduced to below 1.5GB;

– E4B : 8 billion total parameters, 4.5 billion effective parameters, 128K context, MMLU Pro reaches 69.4%, close to the level of the previous generation 27B;

– 26B A4B MoE : 25.2 billion total parameters, only 3.8 billion activation parameters, inference speed close to 4B model, ranked sixth on Arena AI open source leaderboard;

– 31B Dense : 31 billion parameters fully activated, 256K context, ranked third on Arena AI's open source leaderboard.

In terms of performance, it represents a significant improvement over the previous generation Gemma 3 27B. 31B scored 89.2% in the AIME 2026 Math test (compared to 20.8% for the previous generation), LiveCodeBench v6 code test scores jumped from 29.1% to 80.0%, and long document processing MRCR v2 128K improved from 13.5% to 66.4%.

All models support image and video input and over 140 languages, and feature a built-in, toggleable think mode. The E2B and E4B also include an audio encoder with approximately 300 million parameters, supporting speech recognition and translation for up to 30 seconds, and can run completely offline on smartphones, Raspberry Pi, and NVIDIA Jetson Orin Nano.

It's worth noting that the Gemma open-source license has now been fully switched to Apache 2.0, allowing developers to freely modify, distribute, and commercialize it without any user base restrictions. Hugging Face co-founder Clément Delangue described this as a "major milestone."

The Actors Committee of the China Radio and Television Federation issued a statement: AI face-swapping and voiceprint cloning require authorization; derivative works are not exempt from liability.

According to the People's Daily, the Actors Committee of the China Radio and Television Social Organizations Federation issued a stern statement yesterday, taking a concentrated stance against infringements such as AI face-swapping and voiceprint cloning, and demanding comprehensive regulation of the application boundaries of AI technology in the performing arts field.

• Rights Protection : Performing artists enjoy personal rights such as portrait rights, voice rights, and artistic image rights in accordance with the law. Without the formal written authorization of the performer, it is strictly prohibited for any entity to collect, use, synthesize, or disseminate related images, voiceprints, and artistic images.
• "Fan-creation" is not exempt from liability : Infringing content such as AI face-swapping, voice imitation, face-swapping short dramas, and virtual human replicas, even if marked with words such as "non-commercial", "public welfare sharing" or "personal fan-creation", shall still bear full liability for infringement;
• Platform responsibility : Short video, live streaming, film and television distribution and other online platforms must establish and improve the AI ​​performance content authorization verification mechanism, immediately investigate and remove existing infringing works, and retain infringement data to cooperate in evidence collection and tracing.
• Technology platform compliance : AI technology research and development and service platforms must strengthen their pre-approval obligations and verify the authorization qualifications of actors' portraits, audio and other materials uploaded by users;
• Mass rights protection : The Actors Committee of the China Radio and Television Federation will launch a routine online infringement monitoring, notarized evidence collection and mass rights protection campaign to pursue civil, administrative and even criminal liability against individuals, organizations and platforms that fail to fulfill their review obligations for malicious infringement.
• Compliance Advocacy : Support AI technology to empower film and television art innovation under the premise of compliance, and advocate the industry to establish unified and transparent standards and revenue distribution mechanisms for portrait, voiceprint and artistic image authorization.

Guo Rui appointed Chairman and CEO of Zhijie Auto

Zhijie Auto announced yesterday that, following approval by its board of directors, Guo Rui has been appointed as Chairman and CEO, responsible for the company's overall operations and management. This personnel change is seen as a key signal that Zhijie Auto's brand 2.0 strategy is entering a new phase of full implementation.

Dr. Guo Rui graduated from the School of Information Science and Technology at Peking University with a Ph.D. in Science. He has been deeply involved in quantum computing research in his early years. His career spans three major fields: fast-moving consumer goods, consumer electronics, and automobiles. He has worked in brand and marketing roles at Fortune 500 companies such as Procter & Gamble and Huawei.

In August 2021, he joined the newly independent Honor as a founding team member, and successively served as Senior Vice President, President of Brand Marketing System, Group Chief Marketing Officer, and joined the company's EMT (Executive Management Team) and Investment Committee.

Tencent WorkBuddy experienced a sudden login failure; the company apologized and offered 1000 Credits as compensation.

At 09:10 yesterday morning, some domestic users of WorkBuddy and CodeBuddy experienced login issues. The official statement indicated that the team intervened immediately and addressed the problem urgently, and the issue was gradually resolved by 11:45 that same day.

Tencent's Public Relations Director, Zhang Jun, also reposted the apology Weibo post, stating: "We are very sorry that the WorkBuddy login anomaly caused inconvenience and disruption to everyone's work and plans. The issue has now been fully resolved, and we will directly compensate all affected users with 1000 Credits."

Hygon Information officially unveiled its "intrinsic security" technology achievements, aiming to accelerate the training of domestically produced trillion-scale large-scale models.

On April 2, at Hygon Information's 2026 Spring Technology Communication Conference, Hygon officially unveiled several new achievements based on the concept of "intrinsic security" and launched the first annual version of the Hygon DCU software stack, providing computing power support for the training of multiple trillion-parameter large-scale models in China.

Ying Zhiwei, Vice President of Hygon Information, said that the challenges of computing power security in the AI ​​era have exceeded the capabilities of traditional solutions, and breakthroughs must be sought from the lowest level of computing architecture.

On a computing cluster composed of tens of thousands of Hygon CPUs and DCUs, the Hygon team has initially achieved stable operation of 10 trillion model parameters through multiple technologies such as intrinsic security, operator tuning, compilation optimization, and convergence of general computing and general computing. Multiple tests, including MLPerf, have reached the international leading level.

Former Nintendo sales chief: Switch 2 price increase "almost inevitable"

According to GamesRadar, Sean, former head of sales at Nintendo, recently stated on the Kit & Krysta podcast that a price increase for the Switch 2 is "almost inevitable."

Sean points out that multiple factors are currently driving up the cost of gaming hardware: the U.S. government's extensive tariff policies, AI-driven chip demand leading to rising memory prices, and recent sharp fluctuations in oil prices.

Inflation has persisted for a long time, tariffs are not going away anytime soon, and AI-driven chip demand is also pushing up memory prices.

In contrast, Xbox and PlayStation have raised their console prices multiple times. Starting April 2nd this year, Sony raised the price of the PS5 series globally, with the PS5 Pro increasing by $150 (approximately 1032 RMB).

Sean believes that Nintendo is currently taking steps to alleviate the pressure, including pricing digital games lower than physical versions.

Nintendo may be able to postpone the price increase for now, but an eventual price hike is inevitable.

The Huajing S achieves the industry's first L2 mandatory standard ahead of schedule, and all models come standard with Huawei Qiankun Intelligent Driving ADS 4 Pro.

Huajing S announced yesterday that its standard Huawei Qiankun Intelligent Driving ADS 4 Pro has exceeded all the performance requirements of the draft mandatory national standard "Safety Requirements for Intelligent Connected Vehicle Combined Driving Assistance Systems" in advance.

According to reports, the mandatory standard covers 65 field tests, encompassing complex scenarios such as vehicle handling obstacles in tunnels, highway construction detours, roundabout detours, intersections, night driving, and handling vehicles parked at an angle in rain and fog. The standard also requires vehicles to complete all tests at the maximum speed adjustable by the driver.

As the first flagship model created through the deep strategic cooperation between SAIC-GM-Wuling and Huawei Qiankun, the Huajing S proactively participated in preliminary testing and successfully passed all six "hell-level" high-difficulty conditions.

The Huajing S will make its global debut on April 7th at 19:00 at the "Huajing S Leading Technology Appreciation Event". More technical details will be revealed at that time.

JPMorgan Chase CEO Jamie Dimon: AI will bring about a three-and-a-half-day work week, and human lifespan could reach 100 years.

According to CBS News, JPMorgan Chase CEO Jamie Dimon recently stated in an interview that artificial intelligence will fundamentally change the way humans live and work. He predicts that in 30 years, AI could reduce the workweek to three and a half days and help extend human lifespan to 100 years.

I believe that in 30 years, your children may only work three and a half days a week, live to be 100, and many of the diseases that plagued our generation will no longer exist. Life will be better.

He further listed AI's potential in specific fields: "AI will cure cancer, develop better materials, and reduce a large number of traffic accidents."

However, Dimon also acknowledged the potential risks brought by AI. He pointed out that AI-driven productivity gains could impact the job market, especially if the pace of change is too rapid.

The world is becoming more efficient, and AI will drive that process. The risk is that this change may come too quickly.

Regarding the long-term impact, Dimon believes that AI will ultimately create more jobs than it eliminates. He emphasizes that the key lies in businesses and governments preparing in advance, helping workers adapt to change through large-scale retraining and job retraining.

We'll tell employees there are other jobs available. We'll train them and help them transition to other roles. And this approach is effective; it just needs to be implemented on a larger scale.

In the long run, AI will bring enormous benefits to society. If it brings problems in the short term, we must prepare in advance.

With electricity rental starting at 230,000 yuan, the 2026 NIO ES6/EC6 officially launched.

Yesterday, NIO officially launched its 2026 ES6 and EC6 models, with deliveries of both models already underway and expected to arrive at dealerships starting in May. Let's look at the prices:

• The 2026 ES6 starts at RMB 338,000 for the vehicle and RMB 230,000 for the battery rental plan.
• The EC6 vehicle purchase price starts at 358,000 yuan, and the battery rental plan starts at 250,000 yuan.

In terms of configuration, both models introduce the new Moon 3.0 exterior package, add a zero-gravity driver's seat, and come standard with a rear comfort package and two-point magnetic design. They also add a standard driver assistance indicator light, the "little blue light".

The dual-motor combination delivers 150kW at the front and 210kW at the rear. The ES6 CLTC has a range of 630-650km, while the EC6 has a range of 635-655km. Both models have a wheelbase of 2915mm.

Both models are equipped with the world's first automotive-grade 5nm intelligent driving chip, "Shenge NX9031," and the whole-vehicle operating system "SkyOS·Tianshu," which supports point-to-point all-domain navigation assistance and full-model intelligent parking assistance.

In addition, both models are launched with a "Natural Wonderland Edition", whose design is inspired by NIO's global ecosystem co-construction plan Clean Parks. The ES6 Natural Wonderland Edition is priced from NT$350,000, and the EC6 Natural Wonderland Edition is priced from NT$370,000. Battery rental plans are priced from NT$242,000 and NT$262,000 respectively.

The Zhipu GLM-5.1 has been quietly launched, only 2.6 points behind the Claude Opus 4.6.

Yesterday, GLM-5.1 was launched quietly without a press conference or technical report, and was made available to all GLM Coding Plan users (Lite, Pro, Max).

In the official programming ability evaluation, GLM-5.1 scored 45.3 points, only 2.6 points lower than Claude Opus 4.6's 47.9 points . Compared to the previous generation GLM-5's 35.4 points, this single iteration improved by nearly 10 points, an increase of approximately 28%.

The core capabilities of GLM-5.1 have been enhanced in the following aspects:

• The programming benchmark score was 45.3, and the SWE-bench Verified score was 77.8%, which is less than 3 percentage points behind Opus's 4.6.
• Supports a 200K tokens context window and a maximum output length of 131K tokens;
• It performs exceptionally well in agent scenarios such as BrowseComp online retrieval, MCP-Atlas complex tool invocation, and τ²-Bench multi-tool coordinated execution;
• The hallucination rate dropped from 38.6% in the previous generation to 34.1%.

In addition, Zhipu also officially announced the release of its multimodal coding pedestal model, GLM-5V-Turbo, yesterday:

• Natively supports multimodal input such as images, videos, and design drafts; context window is 200k.
• Core capabilities: Understand design drafts/screenshots and directly generate runnable code (front-end replication);
• Deeply compatible with Claude Code and AutoClaw, supporting a complete GUI Agent closed loop;
• It performs exceptionally well on GUI benchmarks such as AndroidWorld and WebVoyager, while maintaining its plain text programming capabilities.

Qwen 3.6-Plus released, supporting millions of contexts.

Alibaba Cloud officially released its new generation large language model Qwen3.6-Plus yesterday, achieving a leapfrog upgrade in programming intelligent agents, multimodal reasoning, and long-term planning, with performance approaching that of the Claude series, and is hailed as the "strongest domestic programming model".

• The capabilities of programming intelligent agents have been greatly enhanced : they have significantly improved their ability to solve code repository-level problems, and support autonomous task decomposition, path planning, and test modifications in scenarios such as front-end development and complex code governance.
• Native Multimodal Awareness : Trained based on native multimodal data, it improves the accuracy of understanding the physical world, parsing complex documents, and video reasoning, and can directly generate front-end code based on design drafts or screenshots;
• Million-level context support : It supports ultra-long context windows of up to 1 million words by default and is optimized for Vibe Coding, greatly reducing the development threshold through natural language commands;
• Ecosystem optimization support : It has been opened through Alibaba Cloud's Bailian API and has been deeply optimized for mainstream intelligent agent frameworks such as OpenClaw, Claude Code, and Qwen Code.

Meituan releases LongCat-AudioDiT audio generation model

Meituan Longmao released the audio generation model LongCat-AudioDiT yesterday, and simultaneously open-sourced two versions, 1B and 3.5B.

According to reports, LongCat-AudioDiT performs diffusion modeling directly in the waveform latent space. The entire process requires only one waveform variational autoencoder (Wav-VAE) and one diffusion transformer (DiT), eliminating the error accumulation caused by multi-stage cascading from the root.

• Training-Inference Alignment : In each step of inference, the hidden variables in the cue area are forcibly reset to true values, thus solving the long-standing problem of timbre drift.
• Adaptive Projection Guiding (APG) : Replacing the traditional classifierless guiding (CFG), it decomposes the guiding signal into orthogonal and parallel components, retaining the beneficial parts and suppressing the degraded parts, thereby improving the timbre similarity while avoiding spectral "oversaturation".

In the Seed benchmark test, LongCat-AudioDiT-3.5B achieved a speaker similarity (SIM) of 0.818 on the Chinese test set (Seed-ZH) and 0.797 on the Chinese difficult sentence set (Seed-Hard), surpassing models such as Seed-TTS, CosyVoice3.5, and MiniMax-Speech, achieving the current state-of-the-art (SOTA) score.

GitHub: https://github.com/meituan-longcat/LongCat-AudioDiT

Hugging Face: https://huggingface.co/meituan-longcat/LongCat-AudioDiT

Lark open-source CLI allows direct control of core workflows.

Lark recently officially open-sourced its command-line tool, Lark CLI (lark-cli), providing AI Agents with a standardized entry point for directly manipulating Lark. After connecting to Lark CLI, AI Agents can directly operate Lark's core business modules, including:

• Messages and Groups: Search, Send, Reply;
• Cloud documents/cloud space: creation, reading, writing, commenting, collaboration, and file management;
• Spreadsheets/Multidimensional Tables: Reading and writing cells, managing views and dashboards;
• Calendar: Check schedules, create meetings, and recommend available time across time zones;
• Video conferencing: Obtain minutes and verbatim transcripts of key notes;
• Email: Search, draft, send, archive (significant enhancements in this update);
• Tasks / Knowledge Base / Contacts / Global Search.

Lark CLI has also made specific optimizations to the AI ​​Agent invocation method: when an error occurs, it explains the repair path to the AI, automatically guides the supplementary authorization when permissions are lacking, and simplifies the token consumption, making the call success rate significantly higher than directly calling the API.

GitHub: github.com/larksuite/cli

Jinmailang announced it will stop using the "handmade" trademark and cease production of related products.

According to Jiemian News, yesterday, Fan Xianguo, founder and chairman of Jinmailang Group, released a video to formally respond to the public controversy caused by the discrepancy between the name and the actual process of his "handmade noodles" product.

In the video, Fan Xianguo stated that the "hand-typed" trademark was registered in 2006 and has been used for nearly 20 years. In order to avoid causing confusion to consumers due to product descriptions, the company has decided to voluntarily stop using the trademark and announced that it will immediately stop producing all products bearing the "hand-typed" trademark starting from the early morning of today (April 2).

The term "handmade" usually evokes images of manual craftsmanship in consumers' minds, but Jinmailang actually uses a mechanized production process, creating a significant discrepancy that has raised suspicions of false advertising.

The wired headphone market has recovered, with sales surging by 20%.

A recent research report from market research firm Circana shows that wired headphones are experiencing an unexpected market recovery:

After five consecutive years of declining sales, including a drop of $42 million in 2024, wired headphones achieved approximately 3% growth last year (about $15 million). In the first six weeks of this year, wired headphone sales surged by 20% year-on-year, reaching a new high in nearly a decade.

Analyst Ben Arnold points out that the factors driving this recovery are multiple. Wired headphones, with an average price of around $13, offer a highly competitive, low-priced alternative to Bluetooth headphones, which averaged around $99 last year.

At the same time, consumers' preference for "retro" products is also strengthening: last year, sales of "CCD" cameras surged by 93%, vinyl records continued to grow, and together with the resurgence of wired headphones, they constitute a wave of "nostalgic consumption".

According to Jimu News, consumers aged 18 to 30 accounted for over 60% of wired headphone sales, with the 18 to 24 age group accounting for 38%, making Generation Z the core driver of this growth.

CASETiFY x Takashi Murakami "FULL BLOOM" Collection Returns

According to NOWRE, the collaboration between CASETiFY and Japanese contemporary artist Takashi Murakami has officially returned. This return is an important part of CASETiFY's 15th anniversary celebration and has special commemorative significance.

The newly launched "Takashi Murakami x CASETiFY: FULL BLOOM" collection features classic characters such as MR. DOB, KAIKAI AND KIKI, and incorporates Murakami's signature FLOWERS pattern. It includes Apple, Samsung, and Google phone cases, as well as accessories with MagSafe functionality and the CASETiFY Travel resilient and shockproof suitcase.

Super Mario Galaxy Movie opens today.

Super Mario Galaxy Movie officially hits theaters today.

This film is a sequel to 2023's "Super Mario Bros. the Movie," telling the story of Mario and Luigi, who, together with Princess Rosata and Yoshi the dinosaur, embark on a galactic adventure across the multiverse to collect Power Stars in order to fight against Bowser Jr., who intends to rescue their father Bowser.

The second season of "Life's Angry" is scheduled to premiere on April 16th.

Netflix has officially announced that the second season of the series "Rage" will premiere on April 16. The new season stars Oscar Isaac, Carey Mulligan, Cailee Spaeny, Charles Melton, Youn Yuh-jung, and Song Kang-ho. Lee Sung-jin will continue to write and create the series, and A24 will remain the production company.

The new season's story focuses on a young couple who work at a country club and witness a heated argument between the owner and his wife, which triggers a series of chain reactions within the club's upper-class circle, owned by a Korean-American billionaire (played by Yoon Yeo-jung).

The first season of "Cheer Up" stars Ali Wong and Steven Yeun. It tells the story of two strangers who become enemies and seek revenge on each other after a car dispute. The show was well-received after its broadcast and won several limited series awards, including Emmy Awards, Golden Globe Awards, Critics' Choice Awards, and Screen Actors Guild Awards.

Falling 2 is set for a North American release in August.

According to the blogger "Hollywood Watcher", the popular 2022 thriller-adventure film "Falling" has announced a sequel, "Falling 2", which is scheduled to be released in North America on August 7 this year.

The sequel stars Harriet Slater, Asema Thomas, and Tom Brittany, with the Spiering brothers, known for "Predestination," directing, and Scott Mann, the original director and screenwriter, continuing to write the script.

In terms of plot, the story focuses on Jax, who is deeply grieving the death of her sister Hunter. She reunites with Hunter's fearless best friend Luce, and the two travel to the famous Kwan Mountains in Thailand for a hike, hoping to heal their inner wounds. However, a sudden landslide traps them on a precarious walkway 3,000 feet above the ground, forcing Jax to confront her fears and fight for survival.

#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.

Just now, Anthropic announced a "ban" on Openclaw.

Many OpenClaw users have quietly received an email in their inbox. The email is short but contains a lot of information: Starting April 4th, Pacific Time, you will no longer be able to use your Claude subscription credit to access third-party tools, including OpenClaw.

Want to continue using it? There are two options.

You can either purchase additional usage bundles (offered at a discount by the official website) or provide your own Claude API Key and pay only for usage. There is compensation: a one-time monthly fee in credits, which will be converted into your monthly subscription fee. For a full refund, please find the link in the email.

Upon hearing the news, OpenClaw founder Peter Steinberger posted a poignant message on social media: "Board member Dave Morin and I tried our best to persuade Anthropic, and the best we could do was postpone it by a week."

"I made it with Claude."

The fact that OpenClaw became popular is itself an absurd story.

Peter Steinberger, the founder, is a legendary figure in iOS development and the creator of PSPDFKit. His coding skills should be unfathomable. However, when he created OpenClaw, he saw himself as a product manager.

The entire OpenClaw codebase, including the backend, frontend, CI/CD, tests, and documentation, is automatically generated by Claude Code. Peter himself is only responsible for describing the requirements in natural language.

I haven't written a single line of code myself.

The underlying technology of this tool is entirely based on Claude: extremely long context, agent tool calls, and multi-step inference planning. From its core to its framework, it is derived from Claude.

However, Anthropic still took action.

In fact, the feud between the two companies had already begun before this ban.

What was OpenClaw originally called? Clawdbot. Anthropic saw it and wasn't happy: "Your name sounds too much like Claude's. Change it." So it was first changed to Moltbot, and finally named OpenClaw.

After the name change, the two sides enjoyed a brief period of peace. Then OpenClaw became an internet sensation.

Managing inboxes, calendars, check-in assistance… users say this thing really works. One account can run ten agents simultaneously, 24/7 without interruption. A $200 Claude Max subscription has been calculated to offer up to $5000 worth of usable computing power.

This essentially means Anthropic is heavily subsidizing every heavy user. And OpenClaw users happen to be among the heaviest users. Putting them on Anthropic's servers is like creating a bloodsucker. So this conflict finally erupted today.

The founder fled to OpenAI

What makes the story even more intriguing is Peter's whereabouts.

At the beginning of the year, OpenAI poached him. What does this mean? OpenClaw, a tool heavily reliant on Claude, now has its founder working for Anthropic's most direct competitor. Anthropic must be experiencing a mix of emotions seeing this situation unfold.

In response to this controversy, Boris Cherny, the head of Anthropic Claude Code, posted four tweets in a row to personally explain the situation.

The first policy is announced directly:

"Starting at 12 PM PT tomorrow, Claude subscriptions will no longer cover the use of third-party tools such as OpenClaw."

Then, following the second point, the reason is explained:

"We have been working hard to meet the growing demand for Claude, and our subscription service was not designed for the usage patterns of third-party tools. Computing power is a resource that we need to manage carefully, and we prioritize our customers who use our own products and APIs."

In layman's terms: You're exploiting us too much; our machines can't handle it. Take care of your own people first.

Article Three addresses compensation:

"Subscribers will receive a one-time credit equal to their monthly fee. If you need more, you can purchase a discounted subscription package now. The link to apply for a full refund will be in your email tomorrow."

The fourth point concludes the discussion, and the art style becomes quite official:

"We want to consciously manage growth in order to serve our customers well in the long term. This change is a step towards that goal."

However, netizens are arguing fiercely. Blogger Yuchen Jin believes: "The Claude plan, priced at $200/month, can actually utilize up to $5,000 worth of computing power, which is an astonishing subsidy. Considering Claude's recent stability issues, and given Anthropic's current GPU computing power constraints, this adjustment may indeed be the right choice."

Blogger Peter Yang stated that Anthropic and OpenAI are now subsidizing heavy users who run multiple agents around the clock with a subscription price of $100-$200 per month, which is exactly the same as the crazy money burning by Uber and Lyft to grab market share back then.

As we all know, after both companies went public, ticket prices nearly doubled within a few years, and it took Uber a full 14 years from its founding to its first profit.

His assessment is that OpenAI and Anthropic are not far from going public, and once their financial data is released, these loss-making subscription plans will inevitably become unsustainable, either leading to price increases or data throttling. He also added insult to injury: "Running local models on the Mac Mini and Mac Studio is looking increasingly appealing now." (doge)

Some are thinking further ahead: "Codex is currently quite generous to third-party tools, given OpenAI's larger GPU reserves. The strategic divergence between the two companies is truly beginning to emerge. We'll see who can hold out the longest."

After wading through the road, the bridge was demolished.

Beyond cost, Anthropic has a deeper purpose in blocking OpenClaw.

Yes, Anthropic is simultaneously promoting Claude Code and Claude Cowork.

OpenClaw's control plane is built on chat applications such as WhatsApp, Discord, and Telegram. Users type in the chat box, and OpenClaw executes the commands. It runs 24/7 on your computer, automatically replying to messages, managing schedules, reading and writing files, and executing code.

But returning to the technical details, OpenClaw itself is not a model; it's a framework, a "shell." Its intelligence largely stems from Claude's modeling capabilities and Claude Code's CLI interface.

The most active users of OpenClaw mostly use Anthropic's API—as long as they can afford the token fee.

Anthropic has clearly recognized this. From its early protest against the name "Clawdbot" citing trademark and copyright issues, to the rapid release of four new features in the past two months—each one precisely targeting a core capability of OpenClaw:

• OpenClaw acts as a text proxy via WhatsApp messages, running on the desktop. Anthropic's response is Dispatch, a persistent thread from phone to desktop.
• OpenClaw uses Discord and Telegram as its control planes. Anthropic released Claude Code Channels, which bridges the two ends using the MCP protocol.
• OpenClaw provides full operating system access permissions, browser control, and application operation. Anthropic's corresponding versions are Computer Use and Claude Code, which share computer access within the same workspace.

But here's the interesting part: OpenClaw's success proves that users have a real and strong demand for this type of agent tool. This is a demand that Anthropic's own products couldn't meet at the time.

In other words, OpenClaw paved the way for Anthropic. Now that the path is clear, Anthropic is conveniently dismantling the bridge.

Under Peter Steinberger's tweet, a netizen commented: "Aside from all the previous experiences, I understand why you don't like them very much."

No one would argue with that statement.

Attached is a translation of the original email:

Hello,
From 12:00 PM Pacific Time / 8:00 PM BST on April 4th, you will no longer be able to use your Claude subscription credits to support third-party harnesses (including OpenClaw). You can still use these tools through your own Claude account, but at an additional cost, i.e., pay-as-you-go, which will be charged separately from your subscription.

Your subscription still covers all Claude products, including Claude Code and Claude Cowork. To continue using Claude to sign in to third-party harbors, please enable additional usage for your account. This policy will first be implemented with OpenClaw on April 4th, but it also applies to all third-party harbors and will soon be rolled out to more tools (learn more).

To facilitate a smoother transition, we will be offering a one-time credit for additional usage, equal to your monthly subscription price. Please claim this credit before April 17th. We will also be offering discounts on pre-ordered additional usage packages, up to 30% off.

We've been working hard to manage usage demands holistically, but these tools have placed significantly more strain on our systems. System capacity is a resource we need to manage carefully, so we must prioritize customers using our core products. You will receive another email from us tomorrow, at which time you can choose to get a refund for your subscription if you wish.

#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.