Pogbagram

The Leapmotor D19 officially launched tonight. When the final price appeared on the screen, this long-awaited large SUV finally unleashed its most lethal move.

The starting price is 219,800 yuan, while the top-of-the-line three-motor performance version is priced at 269,800 yuan.

I don't know if this price intimidated the competitors, but it certainly intimidated me.

In the current automotive market, large SUVs of similar size and configuration generally have a starting price of over 250,000 yuan. Consumers have long been accustomed to paying an extra premium for the market position of traditional luxury brands or leading emerging car manufacturers.

Leapmotor skipped the trial phase and brought a flagship model with impressive capabilities into the competitive 250,000 yuan price range.

At the press conference, Zhu Jiangming, founder of Leapmotor, set a tone for the D19:

The D19 is a top-of-the-line, well-rounded vehicle with no weaknesses. We use the best core components and the best suppliers.

For a full-size SUV, filling the configuration list may not be difficult; the difficulty lies in offering such a restrained price after filling it.

Overwhelm your opponents with parameters

Knowing the starting price of 219,800 yuan and then reading the configuration list of the D19 will create a strong sense of incongruity.

Intelligent cockpit and driver assistance are the most challenging and expensive parts of new car manufacturing. In this area, Leapmotor has not made any compromises in terms of tiered configuration reduction, and the Qualcomm 8797 chip is the basic standard configuration for the entire series.

If you have a greater desire for computing power, the top-of-the-line model will also offer two 8797 Extreme Edition chips, with a total computing power of an astonishing 1280 TOPS.

Matching the computing power is an ample memory pool, with the cockpit having 48GB of RAM and 128GB of storage space, and the driver assistance system being equipped with a separate hardware combination of 64GB and 128GB.

The LiDAR is also standard across the entire range, using the Hesai AT128. When the D19 is delivered, it supports 44 advanced driver assistance functions, including automatic U-turns on narrow roads without markings, 120-meter reversing with tracking, and memory parking across floors.

The "Super Zero" voice assistant, which integrates the Qianwen big data model, also conveniently solves some everyday travel pain points. For example, when the parking space is too narrow to get into the car, you can stand outside the car and use gestures to bring it out.

The powertrain is another highlight of the D19. These days, it's common to equip range-extended vehicles with large batteries, and Leapmotor has made the D19 the range-extended model with the longest pure electric range.

The 80.3 kWh CTC battery pack, coupled with an 800V high-voltage platform, gives this range-extended vehicle a pure electric range of 500 kilometers, and a 15-minute charge can replenish half the battery. It can be driven purely on electric power for daily use, and fuel is used for long distances – a very powerful combination.

The pure electric version is the first to use CATL's 115 kWh super hybrid battery cells, combined with a full-stack 1000V high-voltage architecture. The CLTC has a pure electric range of 720 kilometers, and a 15-minute charge can add 350 kilometers of range.

The 269,800 yuan tri-motor version naturally focuses more on performance, with a combined power of 540kW (about 724 horsepower), a 0-100 km/h acceleration time of only 3.94 seconds, and the ability to perform a compass turn with a minimum radius of 3.6 meters.

Meanwhile, the 50,500 Nm torsional stiffness provides an extremely high ceiling for chassis tuning. The front all-aluminum double wishbone suspension paired with the rear five-link suspension, along with the dual-chamber closed air suspension and CDC continuously variable damping shock absorbers, constitute the hardware foundation of the D19.

The LMC 2.0 fusion control system is responsible for directing this hardware. It can support stable control even with a dual-wheel tire blowout at speeds up to 120 km/h, and can also adjust the suspension stiffness by anticipating road undulations using cameras. Addressing the issue of motion sickness in large vehicles, the system's built-in anti-motion sickness algorithm improves the relief effect by 25.2%.

At the launch event, Cao Li, Vice President of Leapmotor Technology, used the analogy of a kitchen oxygen chamber and a double bed to summarize the lifestyle attributes of this car.

The D19 boasts an interior room utilization rate of up to 88%, offering 6-seater and 7-seater versions. The 7-seater model features a one-touch fold-down rear seat, instantly creating a 1.94-meter flat bed.

The 176-liter electric front trunk comes with a 220V power supply, which can be connected to an extension kit to create an outdoor cooking station; the center armrest box inside the vehicle houses an 8.1-liter automotive-grade refrigerator with heating and cooling, supporting a wide temperature range from -6 degrees Celsius to 50 degrees Celsius.

The most interesting feature is the automotive-grade "Forest Oxygen Chamber," which provides 8 liters of oxygen-rich air per minute and supports both diffusion and nasal inhalation modes. This oxygen supply system is perfect for those who enjoy challenging high-altitude road trips.

In areas with high-frequency contact between passengers and the vehicle, the D19 uses semi-aniline top-grain leather, and the second row features a 120° zero-gravity seat on the right side. With the double-layered soundproof glass and completely flat side windows closed, the 23-speaker Dolby Atmos system kicks in. From auditory to tactile, this car strives to create its own unique sense of luxury.

A high-level game that had to be played

At the press conference, Zhu Jiangming said that the technological foundation supporting the starting price of 219,800 yuan comes from Leapmotor's long-term self-developed system across all domains.

According to publicly available information, 65% of the core components of Leapmotor's vehicles are self-developed and self-manufactured. Around the vehicle assembly plant, Leapmotor has built 17 component factories nearby. From battery packs and electric drive systems to cockpit domain controllers, and then to lighting components and thermal management modules, almost all of these major components are processed on Leapmotor's own production lines.

The highly vertically integrated manufacturing model brings obvious business advantages. From parts rolling off the production line to assembly into the complete vehicle, the intermediate packaging, transportation, and inventory processes are significantly reduced. The profit margins that external suppliers would have needed to obtain are all converted into internal cost reductions by Leapmotor.

By controlling the underlying supply chain, Leapmotor has the confidence to absorb the hardware costs of higher-end models in the 200,000 yuan price range and has complete control over the pricing of the entire vehicle.

This is something Leapmotor has no choice but to do at present.

Based on the 2025 full-year financial report released last month, Leapmotor's overall market size has expanded to a rather astonishing scale. The company delivered nearly 600,000 new vehicles throughout the year, recording a net profit of 540 million yuan and achieving a turnaround from loss to profit. However, when broken down by individual vehicles, Leapmotor's current net profit per vehicle is only around 900 yuan.

In fact, Leapmotor's current sales base heavily relies on its B-series and C-series models, which target the lower-tier market. These high-volume products keep Leapmotor's overall average price per vehicle at just over 100,000 yuan. The 14.5% gross profit margin reported in the financial report represents the maximum profit margin achieved within the existing price range through its fully self-developed R&D model.

In the capital-intensive industry of car manufacturing, such meager profit per vehicle lacks the resilience to withstand drastic market fluctuations. While maintaining a low-price, high-volume strategy can indeed stabilize the basic sales volume, it cannot support the company's investment in advanced technologies for the future.

The procurement of dual Qualcomm 8797 chips, the algorithm evolution of the advanced VLA intelligent driving model, and the research and development of edge AI all require hundreds of millions of yuan in funding. Vehicles priced below 150,000 yuan can hardly sustain these cutting-edge and capital-intensive R&D projects. Leapmotor urgently needs to find a new engine for profit growth.

The D19 is tasked with this crucial mission. Leapmotor urgently needs to leverage this new flagship to raise the price point of its products to the 200,000 to 300,000 yuan range. Only by fully opening up profit margins within this new price range can the company maintain its long-term cash flow for technological research and development.

However, facing the high-end market of around 300,000, it is difficult to break through by simply relying on impressive parameters and solid financial calculations. Leapmotor also needs to establish a convincing value narrative.

Fei Xiang's appearance at the press conference perfectly echoed the brand's rebranding goals. This national-level artist didn't deliberately recite dry specifications; instead, he dedicated a significant portion of his presentation to the actual in-car experience. On stage, he meticulously described the fresh air provided by the automotive-grade oxygen chamber and the quality of music playback through the Dolby Atmos sound system.

Leapmotor hopes to cultivate a high-end image that emphasizes quality of life and riding experience through this approach.

To steal orders in the fiercely competitive large SUV market, Leapmotor must provide consumers with a rational reason to buy. It needs to convince its audience to accept a new consumer discourse centered on technology and practicality, guiding them away from the obsession with brand premiums and towards purchasing decisions driven by core technology and practical experience.

With an impressive configuration and a clear profit target, the D19 stormed into the 250,000 yuan market segment. The final result of this tough battle determined the D19's sales volume and also dictated the upward trajectory of Leapmotor's overall business model.

#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.

As Volkswagen's first dedicated model built on a pure electric platform, the ID.3, which has been on the market for six years, has undergone a major upgrade and has been officially renamed "ID.3 Neo".

Currently, Volkswagen is gradually phasing out the numerical naming system for its pure electric models. For example, the ID.4, which is about to undergo a major redesign, has been confirmed to be renamed ID. Tiguan.

However, given the strong brand equity that the ID.3 has already built in the market, Volkswagen decided to retain the number "3" in its name. The newly added "Neo" suffix is ​​intended to emphasize that this is not a routine mid-cycle refresh, but a comprehensive update from the inside out.

The refreshed ID.3 Neo not only boasts a brand-new exterior design and a larger capacity battery, but also features revisions and upgrades to its interior and interactive experience.

Volkswagen has redesigned the front of the ID.3 Neo to align it visually with the upcoming ID. Polo and ID. Cross, showcasing a new family design language. Compared to the ID. Polo, its overall appearance is more mature and sophisticated.

The most noticeable change to the front of the car is the replacement of the previous generation with a thicker, continuous LED light strip that connects the slightly modified headlights with the black trim strip; the Volkswagen logo has been moved slightly lower, and the front bumper has been replaced with a more aggressively styled air intake.

In addition, the A-pillar, roof, rear spoiler, and tailgate have all abandoned the previous glossy black two-tone finish, and are now uniformly painted in the same color as the body. The side profile remains largely the same as the previous model, but several new wheel designs have been added.

In terms of dimensions, the new car is 23 mm longer, 10 mm lower, and has a 6 mm shorter wheelbase, while the width remains unchanged. Volkswagen says these tweaks give the vehicle a completely new proportion, making it appear longer, lower, and more dynamic than its predecessor.

The interior of the cabin is the core of this upgrade, with Volkswagen making comprehensive corrections to controversial design flaws in earlier models.

The interior features a large amount of higher-quality soft materials, and the dashboard has been redesigned with straighter lines to further highlight the spaciousness of the cabin.

Of all the upgrades, the most notable change is the complete return of physical buttons.

• The much-criticized touch-sensitive "slider" has been replaced by a traditional volume knob and physical seat heating switch;
• The newly designed two-spoke steering wheel not only features an illuminated Volkswagen logo, but also replaces all the previous haptic feedback buttons with physical buttons;
• On the driver's side door, each of the four windows has its own independent control button, completely eliminating the previous cumbersome and counterintuitive design of "two buttons plus a toggle lever".

In terms of intelligent features, the new car is equipped with a 10.25-inch full LCD instrument panel and a 12.9-inch floating central control touchscreen. The vehicle's infotainment system adopts a brand-new UI graphical interface and a more intuitive menu logic, consistent with the upcoming ID. Polo.

It's worth mentioning that the digital dashboard is highly customizable and can even mimic the retro dashboard style of a 1980s Golf, including a nostalgic retro battery gauge.

Consumers can also opt for an augmented reality (AR) head-up display, a panoramic sunroof, front seats with massage and memory functions, and a Harman Kardon premium sound system.

The ID.3 Neo is built on the MEB+ platform, the latest evolution of Volkswagen's MEB platform. This architecture will also be used in the ID. Polo and ID. Cross front-wheel-drive models in the future, while the ID.3 Neo will retain the classic layout of a rear-mounted motor and rear-wheel drive.

The ID.3 Neo, which debuted in the European market, offers three rear-wheel drive pure mobility outputs and battery configurations, resulting in a significant improvement in overall energy efficiency.

• The entry-level model comes with a 50kWh battery, delivering 168 horsepower (140kW / 170PS) and supporting a charging power of 105kW.
• The mid-range version is equipped with a 58kWh battery, boosting power to 188 horsepower (140kW / 190PS), and also supports 105kW charging power.
• The top-of-the-line (High) version is equipped with a 79kWh battery, delivering 228 horsepower (170kW / 231PS) of power. Its maximum fast charging power is increased to 183kW, and this version has a WLTP range of 630 kilometers.

In addition, all new models now feature V2L external discharge capability, which can be used as a mobile power source to supply power to external devices.

Furthermore, thanks to Volkswagen's latest software architecture, the ID.3 Neo not only adds a one-pedal driving mode but also supports smartphone digital key functionality. This new software not only improves the vehicle's underlying performance but also significantly optimizes the user experience.

In terms of safety, the new vehicle is equipped with an enhanced Connected Travel Assist system, which enables smoother semi-autonomous driving and offers optional features such as traffic light recognition and Park Assist Pro with memory function.

In addition, the ID.3 Neo comes standard with active safety systems such as lane keeping assist, forward collision warning, and cornering brake assist.

For consumers seeking ultimate performance, Volkswagen is developing a high-performance successor that is faster than the previous 322-horsepower ID.3 GTX Performance.

It is understood that the car is expected to be officially unveiled later this year and may be branded with the "GTI" badge.

Volkswagen stated that it hopes this comprehensive hardware and software upgrade will enable the ID.3 Neo to better cope with increasingly fierce competition from Chinese new energy vehicle brands and maintain its market competitiveness until the all-electric ID. Golf is launched at the end of this year.

Volkswagen CEO Thomas Schäfer admitted that when he took over Volkswagen in 2022, the brand was gradually losing the core values ​​it represented; and the ID.3 Neo is Volkswagen's first representative work in returning to "the real Volkswagen".

Today, Volkswagen's goal is to once again create cars that are "perfectly integrated into people's lives, trustworthy, and highly practical."

Volkswagen has not yet announced the specific price of the ID.3 Neo in the European market. For reference, the current ID.3 starts at £30,860 in the UK (approximately RMB 285,000).

In the domestic market, the domestically produced ID.3 achieved a remarkable sales record of over 93,800 units in 2024, and once even achieved monthly sales of over 10,000 units. It remains the sales benchmark for joint venture pure electric vehicles to this day.

However, under the intense competition in the domestic new energy vehicle market, the competitive advantage of the domestically produced ID.3 has been gradually weakened. Currently, the domestically produced ID.3, with a starting price of 130,000 yuan, has only sold 8,505 units in the past six months.

If the starting price of the domestically produced ID.3 Neo can be reduced to below 100,000 yuan in the future, it might be able to compete with NIO's Firefly.

#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.

Anthropic's momentum this year has been exceptionally strong.

Not only has its popularity remained high, but its reputation has also continued to rise, firmly establishing it as a top player in the AI ​​industry. Now, almost every morning when we wake up, we see their timely push notifications for new products or features. Over time, the initial excitement has turned into a tacit understanding of "It's you again, as expected."

And just now, the highly anticipated Claude Opus 4.7 was officially released, still the familiar formula, the familiar high-scoring player.

Interestingly, Anthropic was very candid in the announcement, even with a touch of pride: "This is not our most powerful model." The legendary, terrifyingly powerful Claude Mythos Preview is still hidden.

But even though Opus 4.7 isn't the most powerful, it still garnered significant attention. This is because it solved a pain point more important than intelligence: reliability. Not the kind of reliability where it does whatever you say, but the kind where it dares to contradict you when you propose a foolish solution and then corrects its own mistakes.

When reliability becomes a rarer quality than intelligence

Benchmark results show that on the industry-recognized most hardcore SWE-bench Pro, the score of 4.7 jumped from 53.4% ​​in the previous generation to 64.3%, a single-generation upgrade of nearly 11 percentage points, leaving GPT-5.4 (57.7%) and Gemini 3.1 Pro (54.2%) behind.

The CharXiv benchmark for visual inference jumped from 69.1% to 82.1%, corresponding to its newly acquired 2576-pixel long-side recognition capability—more than three times the clarity of its predecessor.

This is not just about "seeing things more clearly." Higher resolution directly leads to a chain reaction of improvements in output quality: the generation of interfaces, the creation of slides, and the layout of documents are all enhanced in terms of detail and precision.

On the MCP-Atlas benchmark, a large-scale evaluation tool, version 4.7 achieved a score of 77.3%, surpassing GPT-5.4's 68.1% and Gemini's 73.9%. In tests conducted by the legal AI platform Harvey, version 4.7 achieved a score of 90.9% on the BigLaw benchmark, correctly distinguishing between "transfer clauses" and "change of control clauses," which have historically been a weakness of cutting-edge models.

However, 4.7 is not entirely in the lead. In the BrowseComp test of Agentic Search, 4.7 actually dropped from 83.7% in the previous generation to 79.3%, and was surpassed by GPT-5.4 (89.3%) and Gemini (85.9%).

This regression is not accidental. An agent that will directly report an error when it encounters missing information and refuses to make up answers will naturally be at a disadvantage when the evaluation standard is based on whether or not an answer is provided.

Beyond the data, a more important question is: what does this "reliability" actually mean in real-world work?

Over the past year, the industry's expectations for large code models have generally remained at the level of "writing a function and finding a bug," but Claude 4.7 has shown a completely different character in early testing.

The head of Replit, a well-known cloud development platform, described it this way: "It would refute me in technical discussions and help me make better decisions. It really felt like a better colleague."

It no longer blindly obeys orders, nor does it fabricate data just to meet deadlines. In tests on the data science platform Hex, version 4.7 directly reports an error when encountering missing data, instead of providing a seemingly reasonable but completely wrong alternative value as in its predecessor. The Hex team even stated bluntly: "4.7 under low resource consumption is equivalent to 4.6 under medium resource consumption."

This trait of "refusing to conform" is precisely what is most scarce in advanced software engineering.

Of course, everything has two sides. A prompt written for the old model might produce unexpected results in version 4.7. Those vague instructions that were previously "understood" by the model will be executed literally in version 4.7. This also means that those who are better at clearly expressing their needs will get better results from version 4.7.

Simply being argumentative isn't enough; an AI that goes on strike when faced with setbacks is also not a good colleague. Another major change in version 4.7 is task resilience.

Previously, large models would often crash and report errors when encountering tool call failures in multi-step tasks. The Notion team's tests found that the tool error rate in version 4.7 was reduced to one-third of the original. More importantly, it can bypass obstacles and continue to complete the task when the toolchain crashes.

When AI stops flattering, true productivity will begin to explode.

In one extreme case published by Anthropic, version 4.7 built a complete Rust text-to-speech engine from scratch without any human intervention—writing neural network models, SIMD kernels, and browser demos, and even feeding the output to a speech recognizer for verification, completing the testing as well.

Vercel, a giant in front-end frameworks, also discovered a behavior never before seen: version 4.7 performs mathematical proofs before writing system-level code. This goes beyond simply writing code and enters the realm of rigorous engineering design.

The cost of hiring AI "senior experts"

To verify its ability to handle details, I set up three front-end interaction scenarios, with only one criterion: whether the details were perfunctory or not, it was immediately obvious.

The first scenario was to create a top-down view of a record player interface, the challenge of which lay in presenting the "metallic sheen" and "breathing halo." Version 4.7 didn't resort to cheap color gradients; instead, it realistically recreated the metallic texture through complex CSS style overlays.

The second scenario was to create an old-fashioned electric fan using only CSS, without JavaScript. Faced with this strictly limited task, some models would secretly violate the rules by using JS, but version 4.7 adhered to them. It created the fan's 3D structure using pure CSS, with smooth transitions between the low, medium, and high speed settings. The perspective and shadow treatment of the base also gave it a realistic feel. It found a very good solution within the rules.

The third scenario involves creating a retro cassette player with the old-fashioned, noisy effect of a videotape. The details of the cassette tape spinning are also present.

Of course, becoming smarter comes at a price. Opus 4.7 is now available on all Claude products and APIs, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry platforms.

The base pricing remains unchanged at $5 per million inputs and $25 per million outputs. However, version 4.7 introduces a brand-new tokenizer that will split the same text into 1.0 to 1.35 times more tokens than before.

In addition, it tends to "think a little longer" during high-intensity tasks, so the actual consumption will almost inevitably increase.

In addition, Anthropic has added a brand new xhigh difficulty level to the existing options. At this level, Claude 4.7 will consume more tokens and spend more time "thinking" when facing complex problems. Claude Code has already raised the default effort level of all packages to xhigh.

Anthropic's actions demonstrate that for real coding tasks, it's better to think things through than to be stingy with resources.

To match this workflow, Claude Code launched two killer features:

/ultrareview (In-depth review): Start a dedicated review session and, like an extremely discerning senior reviewer, thoroughly read through all code changes, accurately identifying deep-seated architectural design flaws and bugs. Pro and Max users can try it for free three times.

Auto Mode extended to Max users: a new permission mode between "granting permissions one by one" and "skipping all permissions". Claude will make decisions autonomously within the scope of your authorization, which can complete long and tedious tasks, and is safer than complete decentralization.

To prevent this "overthinking" AI from maxing out the account balance, the API has also launched a public beta version of the "Task Budgets" feature, allowing developers to explicitly plan Claude's token spending priorities in long tasks.

Of course, 4.7 is not Anthropic's strongest hand.

The more powerful Claude Mythos Preview was just released this month under the name "Project Glasswing" to a small group of companies for cybersecurity research. Mythos has not been publicly released because its cyber attack and defense capabilities are too strong, and Anthropic feels it hasn't figured out how to securely distribute it to everyone yet.

4.7 The system itself has made proactive trade-offs, lowering network attack and defense capabilities during the training phase and incorporating a built-in automatic blocking mechanism to directly block high-risk requests. Security researchers with compliance requirements can apply individually through official channels.

The logic behind not rushing to play your strongest cards and constantly adding new cards to the table is the same. In fact, Anthropic's real moat is the delivery tempo itself.

From February 1st to March 24th this year, in just 52 days, Anthropic updated a total of 74 products, averaging less than one every two days. Cowork, plugins… these actions effectively addressed the pain points of workplace office work.

Today, the Claude ecosystem has long surpassed the simple concept of a "chatbot." For teams eager to deeply integrate AI into their actual workflows, this stable, frequent, and predictable update cycle is the most reassuring reassurance.

Today's release of Claude 4.7 is the latest ballast stone in this chain. And the Mythos Preview will come sooner or later. By then, what we now consider a very strong 4.7 may just be the beginning.

#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.

Would you like to buy a new camera?

For most people, the answer is probably: no.

Smartphones already handle most of my daily recording needs. For professional productivity? Then you're limited to the three major smartphone brands. Focusing on small, high-quality products? Then Ricoh and Fujifilm are sufficient.

Clearly, there is not much room left for innovation in the camera category.

But why is the DJI OSMO Pocket 3 still so hard to find?

For the past three years, the DJI OSMO Pocket 3 has had no rivals—it looks neither like a phone nor a camera, but neither phone manufacturers nor camera manufacturers can ignore it.

Judging from the specifications on paper alone, the DJI OSMO Pocket 3 doesn't even seem to have any purpose, as it could be easily replaced by a phone or camera—but it has become a new species that neither phones nor cameras can kill.

And now, its successor, the DJI OSMO Pocket 4, has arrived.

Is it the Pocket 4 or the Pocket 3 Plus?

It features a single camera with the same lens as its predecessor, and at first glance, it looks exactly like the Pocket 3.

If you only read the parameter table, Pocket 4 doesn't have the sharpness that came with three years of development. But the parameter table has never been Pocket's battleground.

The Pocket 3 has such a great foundation—

The Pocket 3's gimbal stability, portability, and video quality still surpass those of smartphones today, and even rival many cameras. Therefore, the Pocket 3's biggest competitor is itself.

As a long-time Pocket 3 user, in my opinion, what really wears down my patience is not the image quality itself, but rather the small problems that I encounter every time I shoot: zoom and gimbal tilt conflicting, blurry image quality at 2x zoom, and cumbersome switching between photo and video recording…

In these respects, the Pocket 4 is a perfect solution.

The 37-megapixel 1-inch CMOS sensor is the biggest upgrade to the Pocket 4's hardware – high resolution, high dynamic range, and support for 10-bit D-Log are all added at once, ensuring that this device has no obvious shortcomings, whether it is still images or moving videos.

Next, there will be numerous improvements to the user experience in terms of details:

Let's start with the issue of shooting people—DJI has directly added magnetic contacts and a fill light to the Pocket 4. Selfies, indoor shots, and night scenes are the three most common scenarios for Pocket users, and also the most difficult scenarios for any mobile imaging device to handle. With the magnetic fill light, the Pocket 4 finally produces more three-dimensional and better-looking portraits, especially in uneven lighting conditions. The fill light reduces irregular shadows on the face and makes skin tones more beautiful.

Let's talk about the buttons—

The Pocket 4 adds two new buttons: a dedicated zoom button and a camera button.
The previous generation Pocket had limited pixels, and at 2x digital zoom, the image was visibly blurry. With the increased sensor pixels in the Pocket 4, even with cropped zoom, the image remains sharp. Zoom has transformed from a "usable but not very useful" function into a truly reliable composition tool.

The significance of a dedicated shutter button is self-evident—today, photos bearing the DJI Pocket 3 watermark almost stand at the top of the social media hierarchy. Despite only having 9.4 megapixels, a large number of users still take photos with the Pocket 3—most of which are not traditional still images, but Live Photos and three-panel collages—and the most widely circulated plogs on social media today are almost all pioneered by the Pocket. Therefore, this dedicated shutter button carries its own weight. It makes photography no longer an adjunct to video recording, but a core function on par with it.

In addition, there are software upgrades such as in-camera beautification processing and built-in film presets, as well as built-in 107GB of storage space and ultra-high-speed transmission, all of which are designs adopted to improve direct output efficiency. For professional users, support for 10-bit D-Log and 4K 240fps are also highly practical updates.

The shortcomings of the original Pocket 3 have been almost entirely addressed in the Pocket 4.

Of course, there are also regrets. For example, the dedicated telephoto lens that many people have been eagerly anticipating has not arrived as expected. Even with the best image quality, cropping from a wide-angle lens still has a significant gap in creative potential compared to using a dedicated telephoto lens.

Furthermore, we believe there is still more room for improvement in this sensor – for example, with the increased pixel count, it could theoretically support higher resolutions for vertical shooting than the previous 3K. When will DJI update to support 4K vertical shooting in the future? DJI clearly has something else in store.

So, is this the DJI Pocket 4 or the Pocket 3 Plus?

After using it for a while, I feel that the name is the least important issue.

Those moments that made you frown in the Pocket 3 era are almost entirely gone in this generation. No single change is enough to convince you to open your wallet, but when they're combined, things start to change.

Moreover, the Pocket 4 has a more competitive price, which is undoubtedly a killer move in the wave of rising prices for consumer electronics.

The biggest upgrade of Pocket 4 is that it has no weaknesses.

If you're still debating whether to buy the Pocket 4, then I think you already have a better idea by now. I'd like to return to the question posed at the beginning of this review:

Would you like to buy a new camera?

A camera designed for likes

After getting my hands on the Pocket 4, I kept thinking about how to evaluate this new-era camera.

Undoubtedly, the Pocket 4 will continue the success of its predecessor, but it will also face more competitors—it is a huge market.

In less than two years since its launch, the DJI Pocket 3 has sold over 10 million units worldwide, making it the world's most popular portable camera, with revenue approaching 20 billion RMB in 2024 alone.

In 2024, global digital camera shipments reached a near-decade high of 824.7 billion yen, equivalent to approximately 40.3 billion yuan. This figure represents the total shipment value of mainstream camera products (including mirrorless, SLR, and fixed-lens cameras) from companies such as Canon, Nikon, Sony, Fujifilm, and OM System.

In other words, the annual revenue of a single Pocket 3 is roughly equivalent to half of the total annual shipments of all traditional digital camera manufacturers worldwide.

On Xiaohongshu, as of March 2025, Pocket 3-related content had accumulated over 1.2 billion views, over 680,000 posts, and over 8 million discussions —this is a camera that almost everyone dreams of, and they all have mobile phones.

This reminds me of around 2000, when point-and-shoot cameras and digital video cameras once dominated the mobile imaging market—these "clumsy and bulky" professional imaging devices were put into bags and pockets by ordinary people for the first time.

Photography has become a trend that can be embraced anytime, anywhere. Everyone who owns a Nokia or Motorola phone wants to buy a point-and-shoot camera.

For the next decade, annual shipments of point-and-shoot cameras and digital video cameras exceeded 100 million units—until the release of the iPhone 4 in 2010. This was the first iPhone to achieve a significant leap forward in image quality. In the following years, "Shot on iPhone" quickly became mainstream, and smartphones replaced the vast majority of portable cameras.

In 2015, the iPhone officially surpassed Nikon and Canon to become the world's most popular camera device. Also in that year, shipments of compact cameras plummeted to less than 2 million units, officially exiting the market.

Compact cameras and camcorders did indeed revolutionize our photography habits—bringing cameras out and capturing everyday moments. But they didn't change the content distribution process. Essentially, it was a revolution in image miniaturization; the photos they took would eventually become content in albums and on television—they were born for "content."

Smartphones are a further step in this transformation, allowing users to take their phones out and share their social media feeds. The way smartphones have changed the world is by creating new media formats, leading everyone to engage with social media. Thus, the highly integrated nature of "content" and "media" in smartphones naturally made them a replacement for point-and-shoot cameras and digital video cameras.

But the DJI Pocket 3 is a camera designed for "likes".

That rotatable screen is the starting point—shooting long videos on Bilibili in landscape mode, short videos on Douyin in portrait mode, and converting the videos into Live Photos for Xiaohongshu—the same camera, yet adapted to the content needs of different platforms at both the hardware and software levels. Before even starting to shoot, users have already chosen where to post the content and considered what kind of visuals will "get likes".

The DJI Pocket 4 further enhances this capability—it can shoot videos and photos; it offers real-time beautification and film presets; it can store footage without a memory card; even in low light, you can use lighting; want to shoot some amazing professional footage? You can also choose 10-bit D-Log M, 4K 240fps…

It certainly can't replace a mobile phone, but it is definitely a better alternative to a mobile phone—its battery life doesn't consume the phone's battery, its image is more stable than a mobile phone's, its functions are more specialized than a mobile phone's, and its entry barrier is lower than that of a camera—smartphones have too many things to do and too many compromises, but they can't kill a camera like this.

Just as people with feature phones in the 2000s wanted to buy a point-and-shoot camera or camcorder to take photos and videos, people with smartphones today would also want to own a DJI Pocket 4 to shoot TikTok videos and post on Xiaohongshu, right?

After all, who doesn't like getting likes?

This article was written by Zhang Zhanyu and Xiao Qinpeng.

#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.

Claude Opus 4.7 was released late at night, significantly boosting programming capabilities.

Last night, Anthropic officially released the Claude Opus 4.7 flagship model , priced the same as the previous generation Opus 4.6: $5 per million input tokens and $25 per million output tokens. Developers can access it through claude-opus-4-7.

According to reports, Opus 4.7 maintains rigor and consistency in complex, long-running tasks and has the ability to self-verify its output before reporting results. In Cursor's internal evaluation, Opus 4.7 achieved a 70% pass rate on CursorBench, compared to 58% for Opus 4.6.

In terms of multimodal capabilities, Opus 4.7 supports high-resolution image input with a longest side of up to 2,576 pixels (approximately 3.75 million pixels), which is more than three times that of the previous Claude model, opening up space for sophisticated visual scenarios such as computer-controlled intelligent agents and extraction of complex chart data.

In terms of security and alignment, Opus 4.7 enhances the interception of high-risk cybersecurity requests by limiting relevant capabilities during training and incorporating built-in protection layers. To this end, Anthropic has launched a "Cybersecurity Audit Program," which is only available to professionals applying for compliance.

Anthropic also acknowledges that while Opus 4.7 has made progress in honesty and resistance to prompt injection attacks, it has regressed slightly in terms of hazard recommendations for controlled substances, and its overall alignment performance is currently still inferior to Mythos Preview.

OLED iPad Air rumored to arrive next year

According to ETNews, Apple plans to launch a new generation of iPad Air with an OLED display in the first half of next year.

To support this plan, Samsung Display is expected to begin mass production of OLED panels specifically for the iPad Air by the end of this year or January next year. Industry experts predict that the new iPad Air will be released in March or May next year.

It is worth noting that the OLED in the new iPad Air is not the same high-brightness dual-layer stacking solution as the iPad Pro, but a single-layer light-emitting stacking structure with reduced specifications and costs, combined with low-temperature polycrystalline silicon (LTPS) thin-film transistors (TFTs) and a hybrid substrate design.

Lei Jun's official announcement livestream included a 15-hour SU7 long-distance range test: He stated that he wouldn't remain silent just because of negative press.

Yesterday, Lei Jun, founder, chairman and CEO of Xiaomi, announced that he would participate in and livestream the entire long-distance range test of the new Xiaomi SU7 from Beijing to Shanghai at 6:30 this morning, which would last about 15 hours.

He once publicly mentioned that he drove a standard version of the YU7 from Beijing, fully charged, covering 1,300 kilometers on the highway, only needing to recharge once to reach Shanghai, and said he was "very shocked." This statement subsequently sparked controversy.

Lei Jun responded that many of the relevant clips circulating on short video platforms were re-edited by online trolls with the aim of labeling him as engaging in "false marketing."

Lei Jun directly responded to netizens in the comments section: "What should we do? We can't just stay silent because there are negative articles about it, right?" He stated that he would respond to the doubts with concrete actions through this live-streamed test of the new SU7. Lei Jun also revealed that he had been so nervous that he hadn't slept well for several days.

Codex releases major update: background Mac control, image generation, and automatic user preference memory.

Early this morning, OpenAI officially released a major update to Codex, expanding this AI programming assistant for developers to the entire software development lifecycle.

• Background computer use: On macOS, you can independently view the screen, click and input using your own cursor, and run multiple AI agents in the background to process tasks in parallel without interfering with the user's normal operation in other applications;
• By integrating gpt-image-1.5, images can be generated and iterated directly within the same workflow for use in product concept design, front-end interface prototypes, game assets, and other scenarios.
• With the addition of a built-in browser, users can directly annotate and comment on the page to issue precise instructions to the AI ​​agent, currently primarily serving front-end and game development scenarios;
• It supports handling GitHub review comments, running multiple terminal tabs, connecting to remote development environments via SSH, and opening PDF, spreadsheet, and PowerPoint files directly in the sidebar.

In terms of the plugin ecosystem, OpenAI has added more than 90 new plugins, covering skill extensions, application integration, and MCP server access.

In terms of continuous work capabilities, Codex's automation features have been enhanced, supporting the reuse of existing conversation threads to preserve context, and the ability to autonomously schedule future tasks and automatically wake up at a predetermined time to continue executing long-term tasks, with time spans reaching several days or even weeks.

At the same time, Codex also launched a preview of its memory feature, which can remember user preferences, historical error correction records, and information collected over time, thereby improving the speed and quality of subsequent task completion.

Apple releases its 2025 environmental report: recycled materials account for 30% of its products, a record high.

Yesterday, Apple released its annual Environmental Progress Report, announcing that by 2025, 30% of all products shipped will use recycled materials, a record high. CEO Tim Cook stated, "These milestones in protecting our planet demonstrate that ambitious environmental goals can also be a powerful driver of innovation."

The report disclosed specific progress across multiple dimensions:

• Recycled Materials : Our self-developed batteries use 100% recycled cobalt, all magnets are made from 100% recycled rare earth elements, and our self-developed printed circuit boards use 100% recycled gold plating and recycled solder.
• Packaging : Completed a full transition to 100% fiber-based packaging, completely eliminating plastic from product packaging; cumulatively avoided the use of over 15,000 tons of plastic over the past five years, equivalent to approximately 500 million plastic water bottles.
• Renewable energy : Direct suppliers purchased over 20 gigawatts of renewable energy last year, generating over 38 billion kilowatt-hours of electricity; Apple's offices, retail stores, and data centers maintain a 100% renewable power supply.
• Water resources : Apple and its suppliers saved 17 billion gallons of freshwater last year and, through contract projects, replenish more than half of the total freshwater consumed by its global facilities by 2025. All eight of its owned data centers worldwide have received Water Management Alliance (WMA) standard certification.
• Waste : Waste conversion rate reached 75%, with the supply chain transferring over 600,000 tons of waste throughout the year to avoid landfill, and 400 supplier factories participating in the zero waste program; the Fifth Avenue retail store in New York City became Apple's first retail store to receive TRUE zero waste certification.
• Carbon emissions : Compared to 2015, overall greenhouse gas emissions have decreased by more than 60%.

In conjunction with Earth Day, Apple announced that from now until May 16, customers who recycle eligible Apple devices at Apple Stores can enjoy a 10% discount on AirPods or accessories.

The draft national standard for Level 2 assisted driving has been released for public comment: Automatic takeover and system disabling will be implemented if hands-free or eyes-free driving exceeds the permitted time.

The Ministry of Industry and Information Technology (MIIT) yesterday released a draft of the "Safety Requirements for Intelligent Connected Vehicle Combined Driving Assistance Systems" for approval. This is the first mandatory national standard in China for Level 2 assisted driving, and it is recommended that it be officially implemented from January 1, 2027.

Currently, the adoption rate of L2 systems in new energy vehicles has exceeded 50%, but accidents caused by insufficient system capabilities and driver misuse continue to increase. The core purpose of this national standard is to set a mandatory red line for this.

• If both hands are removed from the steering wheel, an alarm must be triggered within a maximum of 10 seconds; if a loss of sight is also detected at the same time, the time is reduced to 5 seconds.
• If eyes are off the road, an alarm must be triggered within 5 seconds, then escalate after another 3 seconds, and finally issue a forced takeover warning within 5 seconds.
• If the driver remains unresponsive, the system will eventually take over the vehicle automatically and pull over to the side of the road.
• If an automatic takeover is triggered once, or if three hands-free/eye-free warnings are accumulated within 30 minutes, the driver assistance system will be forcibly disabled for at least 30 minutes.

In addition, the standard explicitly prohibits the system from changing lanes over solid lines, prohibits driving above the current road speed limit, and requires vehicles to be equipped with Automatic Emergency Braking System (AEBS) and Lane Keeping Assist.

Former DeepSeek researcher Guo Daya joins Seed; ByteDance denies "nearly 100 million yuan annual salary".

According to LatePost, Guo Daya, a former DeepSeek researcher, has joined Seed, a team under ByteDance responsible for the development of large models, as one of the heads of the AI ​​intelligent agent direction. It is conservatively estimated that her total earnings after the four-year stock options vest may reach hundreds of millions of yuan, averaging nearly 100 million yuan per year.

In response, Li Liang, Vice President of Douyin Group, denied the report on social media last night, stating, "This report is untrue, and we have not recently recruited any employees with annual salaries approaching 100 million yuan." He added that if business development is good, it is possible that some Seed technical staff could earn hundreds of millions of yuan in four years.

OpenAI launches GPT-Rosalind, a cutting-edge inference model for biological research.

Early this morning, OpenAI officially released GPT-Rosalind, the first version in OpenAI's life science model series. It has been deeply optimized for modern scientific workflows, with significant improvements in protein and chemical reasoning, genomics analysis, understanding of biochemical knowledge, and access to scientific tools.

In terms of performance, GPT-Rosalind achieved the best results among publicly available scoring models in the BixBench benchmark test; it outperformed GPT-5.4 in 6 out of 11 tasks in LABBench2.

An independent evaluation conducted by OpenAI in collaboration with Dyno Therapeutics shows that the model surpasses 95% of human experts on RNA sequence prediction tasks and reaches approximately 84% of human experts on sequence generation tasks.

Currently, GPT-Rosalind is available as a research preview and is made available to qualified enterprise customers in the United States through OpenAI’s Trusted Access Program. It can be used in ChatGPT, Codex, and API.

Zhao Ming responds to reasons for switching to a car company

According to Phoenix Technology, former Honor CEO Zhao Ming gave his first interview yesterday, explaining his reasons for joining Qianli Technology. Zhao Ming is currently the co-chairman of Qianli Technology, partnered with Yin Qi, an AI entrepreneur and chairman.

Zhao Ming revealed that his acquaintance with Yin Qi at the beginning of this year was purely accidental. They were introduced by a mutual friend and had tea together, after which they communicated on and off for more than half a year. What ultimately impressed him was a proposition posed by Yin Qi: "The commercial and investment closed loop of AI technology is the most difficult problem to solve today, just like the Internet 20 years ago."

Zhao Ming believes that the combination of AI and automobiles is the most imaginative application scenario for the next decade, which is why he decided to join the company. In terms of division of labor, Zhao Ming leads the business operations and commercialization loop, while Yin Qi focuses on technology strategy. The two describe themselves as "powerful together."

Faced with the industry's price war, Zhao Ming's attitude is clear: Qianli Technology will not compete on cost-effectiveness; its goal is to become the world's largest provider of intelligent driving solutions, positioning itself as a B-end "arms dealer" and relying on product reputation to force automakers to recognize the value of its technology.

UnionPay and Ant Group officially announced their collaboration to develop a new payment terminal.

Yesterday, UnionPay and Ant Group officially announced a collaborative ecosystem development partnership, under which they will jointly develop and promote a new type of offline integrated payment terminal.

It is reported that this device will be the first to fully integrate mainstream payment methods such as Alipay tap-to-pay, UnionPay mobile payment, QR code scanning, and card swiping on a single hardware device.

China UnionPay and Ant Group will also collaborate on "trade-in" programs and prize-winning invoices, enabling consumers to participate in national subsidy activities and invoice lotteries through more diverse channels, thus promoting the effective implementation of consumer subsidies.

Google's AI programming tool Antigravity experienced a continuous outage for over 9 hours, with paid users' quotas being locked for up to 7 days.

Recently, Google's AI programming tool Antigravity has experienced continuous outages. On the Google AI developer forum, some users reported that an outage lasting more than 9 hours occurred between April 14 and 15, during which all models, including Gemini 3.1 Pro, Gemini 3.1 Flash, and Claude, which is accessed through Antigravity, were unusable.

Other Pro users who pay $20 per month reported that the advertised 5-hour quota reset period actually turned into a complete lockout of 5 to 7 days, during which time no models could be accessed.

Google's response was to prompt users to upgrade to the $200/month Ultra plan via a pop-up window. However, Ultra users also encountered the same problem of their quotas being silently cut and the service becoming unavailable. One Ultra subscriber warned other users on a forum: "Don't upgrade, Ultra is the same now, quotas are being silently cut, no customer service, no announcements."

Some users have uninstalled Antigravity and switched to Cursor and Claude Code. The message "Goodbye, Antigravity; Hello, Cursor" on the forum has received a lot of likes.

Jensen Huang: If DeepSeek appears first on Huawei's platform, "it will be a terrible consequence for the United States."

In a recent interview with tech podcast host Dwarkesh Patel, Nvidia CEO Jensen Huang warned about the U.S. export control policy on AI chips to China.

In response to Patel's criticism that "selling Nvidia chips to China could help it train AI models with cyberattack capabilities," Jensen Huang bluntly stated, "Your premise is wrong."

He pointed out that Anthropic's Claude model was trained on a fairly average amount of computing power, which is "abundant" in China; China has about 50% of the world's top AI researchers, and the limitation of computing power has instead forced the development of "extremely clever algorithms," and the emergence of DeepSeek is "by no means an insignificant advancement."

He then issued a key warning: "If the first DeepSeek application is released on Huawei, it will be a terrible consequence for the United States."

In his view, extreme export controls would backfire—forcing China’s top open-source models to be deeply optimized specifically for domestic hardware such as Huawei’s. Once these models run best on Huawei’s architecture, the global appeal of Nvidia’s platform will be directly weakened.

If we are forced to leave China, it will accelerate the development of their chip industry and force their entire AI ecosystem to focus on its own internal architecture.

Huang Renxun pointed out that as Chinese AI models spread to the Global South with open standards, the United States may be put on the defensive in the long-term competition for technology standards, and the root cause of all this is the strategic mistake of voluntarily giving up the world's second largest market.

OpenClaw contributor: Building AI workflows on Claude Cowork is like inviting others to exploit you for ten years.

Onur Solmaz, maintainer of OpenClaw and founding engineer of TextCortex, published an article on the X platform yesterday, warning companies against building their core AI workflows on proprietary products of AI giants. He believes this will turn companies into long-term "tenants," facing the risk of being continuously exploited by suppliers. Solmaz did the math:

A subscription to a single large AI lab's premium plan costs approximately $2,400 per year, while a subscription to two labs can reach $4,800; while an ASUS Ascent GX10 local inference device costs approximately $3,000 and can be reused for many years.

He also pointed out that some subscription plans from AI giants are currently being sold at a loss, and price increases are almost inevitable in the future.

What you're really saying is—exploit me as much as possible for the next 10 years, because replacing you would be too painful.

He suggested that companies switch to open-source AI agent frameworks such as OpenClaw and Open Agents to retain the freedom to replace the underlying model or switch local hardware at any time.

DJI releases the Osmo Pocket 4 handheld gimbal camera, starting at 2999 yuan.

Yesterday, DJI officially released the Osmo Pocket 4 gimbal camera. Let's take a look at the price:

• The standard bundle is 2999 yuan (the initial price of the Pocket 3 was 3499 yuan).
• The all-in-one kit is priced at 3799 yuan (including a magnetic fill light, a wide-angle lens, and a DJI Mic 3 kit).

In terms of specifications, the Osmo Pocket 4 features a new 1-inch 27MP CMOS sensor (with an upgraded 14-stop dynamic range), supporting up to 4K/240fps recording; it also supports 2x lossless zoom and up to 4x zoom, and has a larger 1545mAh battery; it includes 107GB of internal storage and adds USB 3.1 and WiFi 6 connectivity.

Related Reading: First Look | DJI Pocket 4 Review: A Camera Built for Likes

Honor launches its first "shrimp-raising laptop," the MagicBook 14/16 2026, starting at 6999 yuan for the first time.

Honor officially launched the MagicBook 14/16 2026 laptops yesterday, focusing on long battery life and featuring the new third-generation Intel Core Ultra processors, offering three processor configurations: Ultra 5 336H, Ultra 5 338H, and Ultra X7 358H.

• Both models are equipped with a 92Wh battery. According to data from Honor Labs, the MagicBook 14 has a battery life of 15 hours and 36 minutes, while the MagicBook 16 has a battery life of 15 hours and 18 minutes.
• The MagicBook 14 features a 2880×1800 120Hz IPS matte screen with a peak brightness of 430 nits;
• The MagicBook 16 features a 2560×1600 180Hz screen with a peak brightness of 500 nits, and both models support 180° one-handed opening and closing.
• The MagicBook 14 weighs 1.39kg, and the MagicBook 16 weighs 1.69kg.

In terms of pricing, the MagicBook 14 2026 Ultra 5 336H + 32GB + 1TB version has a suggested retail price of 7799 yuan and an initial sale price of 6999 yuan; the Ultra X7 358H + 32GB + 1TB version has a suggested retail price of 9599 yuan and an initial sale price of 8799 yuan.

The OPPO Pad 5 Pro supports PC-level WPS and is scheduled for release on April 21st.

Yesterday, OPPO released a graphic summarizing the core specifications of its flagship tablet, the Pad 5 Pro. The new phone will be officially launched on April 21st at 19:00.

The OPPO Pad 5 Pro is positioned as a productivity device that functions as both a tablet and a computer. It supports PC-level WPS, panoramic free-window mode, CapCut Pro, computer screen extension, and keyboard and mouse sharing, and also features the new OPPO File Manager. Specific configuration details are summarized below:

• Equipped with the fifth-generation Qualcomm Snapdragon 8 Ultra + Tide Engine;
• The 13.2-inch 3.4K (3392×2400) screen has a peak brightness of 1000 nits, a contrast ratio of 1500:1, and is certified by TÜV Rheinland Smart Eye Protection 5.0.
• It has a built-in 13380mAh battery and supports 67W fast charging;
• It is available in three colors: Mocha Brown, Monet Purple, and Dawn Gold, and four storage options: 8GB+256GB, 12GB+256GB, 12GB+512GB, and 16GB+512GB.

InStone Luna releases sample images with dual-camera telephoto Dolby Vision lenses.

Yesterday, InStone CEO Liu Jingkang released 4K Dolby Vision sample images of the company's new Luna camera. The first batch of content covers three dimensions: night scene, portrait, and 1-6x zoom, showcasing the end-to-end image quality performance of this gimbal camera in complex lighting environments.

In the night scene samples, Luna excels in highlight suppression and shadow noise control, with official images showcasing a harmonious blend of blue-toned shadow detail and high-contrast lighting in the snow-capped mountains. For portraits, Luna achieves natural bokeh and accurate skin tone reproduction.

In terms of performance across the entire focal length range, Luna supports 1–6x Dolby Vision recording throughout, ensuring seamless image quality.

Notably, this is the first time a gimbal camera product has achieved full-link Dolby Vision connectivity across all focal lengths—from shooting to output, HDR specifications are no longer limited by the device and lens focal length. Liu Jingkang stated at the launch, "6x zoom and 4K are not the end goal; we hope to give users more choices."

Priced at 298.9 yuan, Dreame launches the NB01 electric toothbrush, featuring a buttonless design.

Yesterday, Yu Hao, CEO of Dreame Technology, announced on Weibo that Dreame Dental has officially launched the NB01 keyless smart sonic electric toothbrush, which will be first launched on JD.com.

According to reports, the NB01 is equipped with the "Fit Clean" keyless sensing system, which enables fully automatic sensing control that allows users to "pick up and brush, and stop when they leave their mouths," completely eliminating the need for manual button presses.

• The high-frequency sonic motor operates at 36,000 times per minute, and the manufacturer claims it can complete a quick cleaning in 15 seconds.
• The plaque removal rate reached 98.3%;
• High-precision intelligent pressure control can automatically adjust the power output according to the brushing force;
• The dual brush head configuration caters to both deep cleaning and gentle care scenarios.
• The entire device is IPX7 waterproof certified.

In terms of price, JD.com shows that the NB01 is priced at 298.9 yuan and is available in two colors: "Pearl White" and "Obsidian Black". There is also a "couple set" priced at 498 yuan.

Qwen3.6-35B-A3B Open Source

Tongyi Labs yesterday open-sourced the multimodal model Qwen3.6-35B-A3B, which has a total of 35 billion parameters and activates only 3 billion parameters in each inference.

• Extremely low inference cost: with only 3B activation parameters, it can outperform dense models with 27 billion parameters on programming benchmarks;
• Improved programming capabilities: Significant progress compared to the previous generation Qwen3.5-35B-A3B on benchmarks such as SWE-bench and Terminal-Bench;
• Native multimodal: Supports "thinking/non-thinking" dual modes, covering text, images, documents and spatial intelligence tasks; scores 92.0 on RefCOCO and 50.8 on ODINW13.

MiniMax releases its first cloud sandbox agent, MaxHermes.

Yesterday, MiniMax officially launched MaxHermes, the world's first cloud-based sandbox AI agent.

MaxHermes is built on the Hermes Agent, and its core capability lies in its learning closed-loop mechanism: after completing a complex task, the system automatically extracts reusable skills and saves them as independent documents, which are then loaded on demand and continuously iterated.

Combining cross-session persistent memory, naturally defined timed tasks, and parallel operation of multiple sub-agents, MaxHermes is positioned as a long-term online, continuously evolving AI agent. The underlying model uses the newly released M2.7, one of the most widely used models in the Hermes Agent suite.

Currently, MaxHermes has been integrated with mainstream IM channels such as Lark, DingTalk, and WeChat Work, and supports Token Plan for fee deduction. MiniMax stated that it will integrate with Skillhub in the future and support MaxClaw users to migrate existing configurations with one click.

Alibaba's ATH (AI-Powered Horizontal Transformation) platform launched Happy Oyster, a world model that generates interactive "living worlds" in real time.

Alibaba's ATH team officially launched the world model product Happy Oyster yesterday and opened early access applications.

Happy Oyster is positioned as an open-world model, designed for real-time world creation and interaction scenarios. It adopts a native multimodal architecture, supporting multimodal understanding and joint audio and video generation. The model can continuously listen to and respond to user commands during the generation process, with the scene reacting in real time and evolving continuously with the user's ongoing input.

• Director Mode : Users become real-time directors, able to switch shots, direct characters, or change the story's direction at any time via text, voice, or images. Generated content can be up to 3 minutes long with a maximum resolution of 720p.
• Wandering Mode : Users freely explore an infinitely expandable world in first-person perspective. Object positions are stable, the environment persists, and the viewpoint and lighting respond in real-time to movement. It supports continuing to generate beyond the boundaries of the original scene. Maximum duration is 1 minute, 480p resolution.

Both modes support multimodal input, and the output is both audio and video.

Tencent's Hunyuan releases 3D World Model 2.0

Yesterday, Tencent released and open-sourced HY-World 2.0, which supports multimodal input including text, images, and videos. It can directly output 3D assets in 3DGS/Mesh format that can be imported into mainstream game engines such as Unity and UE, and seamlessly integrate with existing game workflows.

Unlike mainstream world models such as Google Genie 3, which can only generate video files, HY-World 2.0 generates content that can be directly edited and imported into game development or embody simulation engines, making it truly "playable and usable".

Users can input a sentence or an image to generate roamable 3D spaces in various styles, such as realistic, comic book, and game-like. The built-in character mode supports physical collision interaction. The model also supports digital twin reconstruction; by inputting real-world video or multi-view images, a high-precision digital twin scene can be constructed, generated once and permanently reusable.

In terms of technical architecture, HY-World 2.0 takes 3D generation as the main axis and connects four modules: panoramic image generation (HY-Pano 2.0), trajectory planning (WorldNav), world expansion (HY-WorldStereo), and world synthesis (HY-WorldMirror 2.0) to achieve end-to-end generation from multimodal input to a complete 3D world.

GitHub: github.com/Tencent-Hunyuan/HY-World-2.0

LingBot-Map, an open-source streaming reconstruction model for Antminer.

Ant Financial announced yesterday that it has open-sourced its streaming 3D reconstruction model, LingBot-Map. This model requires only a single ordinary RGB camera to estimate camera pose and reconstruct the 3D structure of a scene in real time during video capture, providing continuous and stable spatial perception capabilities for applications such as robotics, autonomous driving, and AR glasses.

LingBot-Map adopts a streaming architecture, receiving video frames while outputting localization results and 3D structures in real time. It supports real-time inference at approximately 20 FPS and can run continuously in long videos of over 10,000 frames with almost no loss of accuracy, balancing the three key indicators of accuracy, speed, and long-term stability.

LingBot-Map outperforms existing streaming methods across multiple mainstream international benchmarks.

StepAudio 2.5 TTS Released by StepAudio

StepAudio 2.5 TTS, a new generation speech generation model, was released yesterday by StepAudio, which reconstructs the control method of speech synthesis around three core capabilities:

• Global context control: Natural language is used to set the emotional tone, character state, and scene atmosphere for the entire speech, replacing the traditional tag combination, resulting in more unified and precise expression;
• Contextual control in the text: Intervening in every detail of the voice, covering emotional intensity, rhythm, stress, pauses, breathing, and even the character's psychological activities and the flow of subtext;
• Zero-sample replication and full timbre control: Any timbre can be used directly without retraining, and a complete "voice character profile" can be built for each timbre, realizing systematic customization from voiceprint to personality.

Leap Star summarizes the positioning of this model as "more than just reproducing sound, but creating expression." The new model covers multiple application scenarios such as audiobooks, film and television dubbing, and intelligent voice interaction.

"Intangible Cultural Heritage + Coffee": Starbucks Launches New Lingnan Cultural Tourism Experience in Guangzhou

Starbucks launched a Lingnan intangible cultural heritage tourism experience event in Guangzhou yesterday, with 11 stores in Guangzhou and Foshan, including Yongqingfang, Shamian, Guangzhou Tower, and Foshan Lingnan Tiandi, integrating traditional Lingnan elements such as lion dance, Cantonese opera, and arcade buildings into the coffee scene.

In terms of products, more than 400 designated stores in Guangdong and Foshan will launch two Lingnan-limited special drinks: Sea Salt Lemon Ice Shake Concentrate and Mango Coconut Symphony Frappuccino. At the same time, lion-themed refrigerator magnets, blind boxes and other peripheral products will also be launched.

The Yongqingfang store is the sixth Starbucks intangible cultural heritage concept store in China, with Cantonese opera mask rattan installations and a modern bar forming the main color scheme of the space. Starbucks stated that this event is an extension of its "1,000 stores, 1,000 faces" strategy, aiming to create a replicable model of "intangible cultural heritage + coffee".

Jasmine Milk White crosses over into the coffee market; Honduras SOE unveils two new products.

Jasmine Milk Tea announced yesterday the launch of its first new coffee series: 400-Serving Thai Milk Tea and Needle King Apple Latte. Both products use 100% Arabica coffee beans from Honduran estates and are presented in SOE (Single Origin Espresso) format.

According to reports, this batch of coffee beans was fermented in whiskey barrels for 30 to 40 days and then medium-roasted. After extraction, it has a triple aroma of Baileys, champagne, and vanilla.

The brand uses the fermentation philosophy of "you reap what you sow" to run through the theme of this new product launch, and quotes the lyrics of the song "When I Become Myself" that "If you have made up your mind, I wish you will sprout like a seed and reap what you sow, create a temperature difference with the world, and bloom an extraordinary flower", with transformation and self-growth as the core of the brand narrative.

Traffic management authorities: It is illegal for green license plates for new energy vehicles to turn white.

According to CCTV News, recently, there have been reports of new energy vehicle license plates being deliberately altered in various parts of China. Some car owners have used methods such as soaking in chemical solutions or exposing their license plates to special lights to fade the green gradient background of the new energy vehicle license plates to white background with black lettering, and have been spreading related "license plate fading tutorials" on social media platforms.

In response, the traffic management department of the public security bureau clarified yesterday that my country has not yet issued any new types of license plates for new energy vehicles, and the aforementioned behavior is illegal.

Zhu Jian'an, deputy director of the Publicity and Education Department of the Road Traffic Safety Research Center of the Ministry of Public Security, pointed out that Article 95 of the Road Traffic Safety Law of the People's Republic of China clearly stipulates that anyone who intentionally obscures, damages, or fails to install license plates in accordance with regulations will be given a warning or fined between 20 and 200 yuan in accordance with Article 90.

For more serious offenses, according to Article 9 of the "Administrative Measures for Demerit Points of Road Traffic Safety Violations," driving a vehicle without license plates or with license plates intentionally obscured or damaged will result in 9 demerit points being deducted from the driver's license.

The public security traffic management department emphasized that the style, color and specifications of motor vehicle license plates are uniformly stipulated by the state. Each license plate corresponds to specific vehicle and owner information and is not a decoration that can be changed at will.

"Four Crossings" is scheduled for release on June 26.

According to Sina Movies, the film "Four Crossings of the Chishui River" has officially released its poster, announcing its release date of June 26. The film is set against the backdrop of the classic battle of the Red Army's "Four Crossings of the Chishui River" during the Long March, striving to recreate the legendary epic of "winning with fewer troops against a larger force."

Michael Jackson biopic "Road to Stardom" begins pre-sales and will be released next Friday.

According to blogger "Hollywood Watcher", the Michael Jackson biopic "Michael Jackson: Road to Stardom" has announced the start of pre-sales in mainland China and is scheduled for official release on April 24, simultaneously with North America.

The film stars Jafar Jackson, Jackson's nephew, as MJ, with Colman Domingo, Nia Long, Laura Harrier, Myles Turner, and others in supporting roles. It is directed by Antoine Fuqua, who previously directed "The Equalizer" and "Training Day".

The film has received support from the Jackson Estate. According to the description, "The story will deeply explore the complex man who later became the King of Pop, and will recreate Jackson's most iconic performances, providing a deeper understanding of his artistic journey and personal life."

"Peppa Pig's Perfect Vacation" releases "China Adventure" trailer, set for summer release.

According to Weibo Movies, the first 3D animated feature film of "Peppa Pig," titled "Peppa Pig: Perfect Vacation," released a "China Adventure" trailer yesterday.

In the trailer, 3D versions of Peppa and her brother George announce their upcoming vacation in China, showcasing a summer itinerary. The film, produced by the original British team, is scheduled for release in China in the summer of 2026.

#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.