大模型 · 739

The Xiaomi MiMO-v3-Pro score is suspected to have been leaked, and the SW-Bench Pro reached 72.8 or close to the top overseas closed source model

According to Twitter news, Max For AI published an article that revealed that a Benchmark screenshot suspected to be Xiaomi's next big model, MiMO-v3-Pro, was circulating in the community. The screenshot shows that the model focuses on coding agents and general agent scenarios, and some test results are close to top overseas models such as Claude Opus and GPT. According to the suspected screenshot, MIMO-v3-Pro scored 72.8 points on the SW-Bench Pro, which is higher than the 67.9 points of GLM 5.3 and the 65.8 points of Kimi K3, which is less than 3 points different from Claude Opus 5's 74.6 and GPT-5.6 Sol Max's 75.4 points; Terminal-Bench 2.0 scored 70.6 points, which is also close to Claude Opus 5's 72.0 points and 73.5 points for GPT-5.6 Sol Max. Furthermore, it scored 76.4 points on the bt3-bench compared to the GPT-5.6 Sol Max with 78.8 points. If the above results are finally officially confirmed and replicated in the official version, the MIMO-v3-Pro may enter the first tier of the world's top models. However, at present, the authenticity and testing conditions of this Benchmark screenshot have not been officially confirmed by Xiaomi, and the relevant data should still be regarded as unconfirmed breaking news. It is worth noting that the latest MiMO flagships officially unveiled by Xiaomi are MiMO-v2-Pro and MiMO-v2.5-Pro, so whether V3-Pro exists and when it will be released is yet to be further disclosed by the official authorities.

1d ago

Whoever sings down Anthropic may be disappointed

Author: Alan Walker, Silicon Valley Original title: Is Anthropic's Growth Slowing Down? Source of controversy. Claude Code ARR tracking chart produced by TickerTrends. The latest data is $15.12 billion for the week of August 10, 2026, accounting for 21.9% of Anthropic's total ARR. Please note: This is an estimate from a third party agency and is not an official disclosure of Anthropic. The first section below explains how important this difference is. Alan Walker from Silicon Valley made an appointment for dinner in Hong Kong. After some hard work, he discovered that this picture had been retweeted more than 30 times, and the matching statement was similar — “Anthropic's growth has leveled off; 2 trillion dollars is a bubble.” Alan saved the image, zoomed it in, and looked at it again. The problem isn't in this picture. This picture is very well done, and the data is probably done seriously. The problem is that almost everyone who retweeted it was using it to answer a question it couldn't answer at all. 01 Let's first figure out who made this picture, there is a Claude icon in the upper left corner. The color scheme is Claude's familiar orange. At first glance, it looks like an official product. It's not. The author of this picture is TickerTrends and has his name written in the upper right corner. It is a third-party data tracking agency that uses various external signals (application data, payment panels, recruitment, channel caliber, etc.) to estimate the revenue of an unlisted company. The line in the picture is written very honestly: “tracked allocation” -- the percentage of allocations that have been tracked. Let's be clear: Anthropic has never publicly disclosed Claude Code's individual ARR numbers, not once. Every point on this curve has been estimated by an outsider. For example, this is like someone using “long queues at the entrance of a restaurant every day” to estimate its turnover and then draw a beautiful weekly curve. The length of the team does correlate with turnover, but in the middle there is turnover rate, customer unit price, takeout ratio, private room business — you see that the team is three short weeks, and the kitchen is probably being renovated in those three weeks. What is more important is the caliber itself. ARR's algorithm is “revenue for the most recent period times 12.” Enterprise software contracts are not executed evenly every day; they are signed batch by batch. Big orders signed at the end of a quarter will jump a week's curve by a large margin; if the next quarter's big orders aren't signed, the curve will go sideways. Weekly ARR tracking is extremely insensitive to this kind of blocky landing—it will paint the “pace of signing” as a “change in demand.” In a nutshell, what you have in your hand is an unofficial weekly map estimated by an outsider, with a very blunt caliber. Judging by the weight of the “bubble” under it is tantamount to using body temperature to measure blood pressure. 02 I hit myself in the face on this picture. I haven't seen anyone mention it, but it's the most interesting part of the whole thing. The picture shows two numbers: Claude Code is $15.12 billion, or 21.9% of Anthropic's total ARR. By dividing: calculate 15.12 billion ÷ 21.9% = about $69 billion. This is Anthropic's total ARR for the week ending August 10, implied by this image. The official caliber figures reported by Bloomberg, Reuters, and CNBC on August 17 were — $65 billion at the end of July. Clear: This chart, which is being used to prove “slowing growth,” its own implied total number of companies is 4 billion US dollars higher than the official figure ten days ago. Further 10 days until today, if the trend continues, more than 70 billion is a reasonable estimate (this sentence is an inference, not data). In one sentence, people who retweeted only read the number 151.2 and the height of the column, skipping the 21.9% next to it. And that 21.9% said: This company went a step further when everyone shouted “it's slowing down.” I only believe in the two numbers on the same picture that is beneficial to my opinion; this is not called analysis. 03 You are looking at the picture below. The money in the picture above has the upper and lower two pieces. Above is the absolute amount (how many billion dollars), and below is the percentage change (how much more than a percent increase from four weeks ago). The vast majority of people's reasoning is: below...

1d agoWendy#Anthropic #ARR #IPOs #MiniMax

Millions of dollars are rushing into the market, but some are in a hurry to exit: Pharos's high-interest treasury causes a “view of time” collision

Article: Sanqing, Foresight NewsSharos Network joined forces with Vault infrastructure agreement R25 and credit asset management agency Axil to launch Axil Prime Credit Vault (APC), an institutional consumer credit RWA wealth management product issued by Pharos on July 15. The products were launched simultaneously with Binance Wallet, TopNod, OKX Wallet, Bitget Wallet, and KuCoin Wallet, with a total fundraising limit of 100 million USDC, with a target annualization of about 14.3%. As of the closing of the deposit window, a total of $45.39 million had been deposited. This year, there have been frequent security explosions in Web3 on-chain strategies. User funds are looking for new stable income sources, and project parties are also there. Binance Wallet is now offering an additional $300,000 PROS as an incentive to explore RWA Vault's market space, causing the Vault to generate a lot of discussion in the market. The launch time coincided with the redemption period of the Pharos TGE pre-deposit campaign. The previous treasury required the submission of a redemption application about half a month before the end of the lockdown period, stop accruing interest on July 20, and complete the redemption within 7 days. Users accustomed to DeFi T+0 looked back and found that they couldn't help but missed the redemption period and began to question the redemption time and asset safety. R25 and Axil then held an AMA at Binance Square. Well-known KOLs such as Haotian and Tianqing participated in discussions, detailing the differences between RWA assets and DeFi Vault, the role of fund managers (Curators), why consumer credit is worth allocating, and risk management methods from pre-investment to post-investment. In complex asset logic and mixed social media discussions, some users put in one million funds on the last day, while others sought early redemptions from the project party. On July 23, Pharos issued an announcement: Users who submitted applications on time in the previous issue have received all principal and interest, breaking the “financial security” concerns; funds that missed the window will automatically be carried forward to the next three-month cycle according to the treasury's preset rules, and interest will continue to be accrued at 14% USDC per annum. The controversy revealed more important issues than the redemption itself. Although the RWA TVL has exceeded $38 billion, non-institutional chain users are clearly dissatisfied when investing in RWA products. Institution-driven, stable, and high interest rates, but often require longer lockdown periods and complex understanding costs. From DeFi to RWA, is the market really ready? High yield, low threshold, and high liquidity. BlackRock's “impossible triangle” of RWA's BUIDL threshold is $5 million. It is only open to qualified buyers, yet it can be redeemed almost instantly through the stablecoin channel; the APC threshold is so low that ordinary users can buy it at will; instead, it must be locked for three months. Liquidity has never been determined by how high or low the threshold is, but rather how quickly the underlying assets can be realized. The bottom layer of BUIDL is US treasury bonds, and the world's deepest secondary market can take over at any time; the bottom layer of APC is hundreds of thousands of emerging market consumer loans, and few people are ready to buy large amounts of capital at any time. This has formed a triangle that RWA cannot bypass at this stage: high yield, low threshold, and high liquidity; the three can only take two. For example, Franklin Templeton's BENJI starts at $20 (low threshold) and supports daily redemption (high liquidity), and the annualization is only 3% to 5%; if you want double-digit returns, you have to accept non-standard assets and a lock-up period. This is the liquidity premium. A significant portion of the excess income is the consideration for abandoning liquidity. APC, on the other hand, is a combination of high returns and a low threshold, and the cost is liquidity. There is nothing wrong with this trade-off itself; it also explains the full source of this controversy. Retail investors have obtained assets that were originally only open to institutions, and they have also taken over the agency's time rules that focus on long-term matching. The period of use of institutional funds is scheduled before investment, and the lockdown period is a predictable cost; private equity credit and closed-end funds already have redemption restrictions. However, most ordinary users on the chain are not the same; most of the latter's first appeal is to go in and out. So the current “retail” RWA is mostly just distribution-side retailing, to be precise. Web3 wallets and low initial investment amounts have contributed to a low threshold, but the liquidity structure is still designed according to institutional logic. Having understood this triangle, the remaining questions became specific: why must the liquidity side be sacrificed, a high income of 14.3%...

2d agoForesight News#WEB3

DingTalk launches AI office app QwenNote, hardware QwenNote A2 revealed

Comparative news, according to the “Reading Best” report, DingTalk is promoting a new AI office application QwenNote (Listen to a Thousand Questions). The application is positioned as an AI portable assistant. It integrates real-time voice transcription, summary, and translation through a combination of software and hardware, and is deeply integrated with AI Agent to embed Agent capabilities into voice input and promote a shift from simple recording to automated execution. The application supports real-time transcription and Chinese-English bilingual recognition and language switching, can generate structured meeting minutes, outlines, and to-do tasks, and has built-in AI Q&A and shortcuts based on memorized materials. QwenNote provides a voice memo function. You need to scan the code to connect to the recording device. Press and hold the button on the back of the device to record inspiration. After the recording is completed, it automatically files, generates a title and brief summary, and marks the time. The product also has an incognito protection mode. After opening, the original audio will be physically deleted, and only the transcribed text will be kept to suit confidential scenarios. The supporting hardware QwenNote A2 has been introduced in the app. It also belongs to the Thousand Question Listening hardware ecosystem. The ecosystem also includes DingTalk A1, DingTalk A1 Pro, Cleer H1, etc., and users can scan the code to complete the binding. According to the report, DingTalk hopes to use the integration of software and hardware to complete offline voice collection portals to form a closed loop of live recording, real-time bilingual transcription, AI minutes Q&A, and DingTalk organization collaboration. The listening material can be synchronized to DingTalk AI to listen to and support personal private isolation. The software side continues to embed large models in scenarios such as documents, meetings, and IM, while the hardware side expands the Thousand Question Listening product line. More official release information has not been widely publicly retrieved for related hardware.

2d ago

Yushu Technology announced the promotion of “self-evolution of physical AI robots”

Comparatively, Wang Xingxing, founder and chairman of Yushu Technology, announced at the main forum of the 2026 World Robotics Conference that Yushu Technology promotes “physical AI robot self-evolution”, which is driven by large AI models, automatically searches for papers, writes control codes, verifies and deploys real machine tests in a simulation environment, and then uses AI and humans to jointly score feedback to form a closed loop iteration. Once the system is in operation, development efficiency and iteration speed can be greatly improved. Wang Xingxing mentioned that in the next decade, against the backdrop of the AI explosion, robots will evolve faster than expected. “This is a new starting point.” The critical flashpoint of the industry in the future will be when users bring robots to any unfamiliar environment and can complete about 80% of daily tasks through language or text instructions alone. This node is expected to arrive in 2-3 years at the earliest and 5-10 years at the slow speed. (every time)

2d ago

Yu Shu Wang Xing: It may take 2 to 10 years for ChatGPT to become intelligent

Comparatively, at the 2026 World Robotics Conference, Wang Xingxing, founder and chairman of Yushu Technology, delivered a keynote speech entitled “From Exhibits to Products: The Next Ten Years of the Humanoid Robot Industry”. He said that in recent years, the company has continued to promote robots into our lives, including implementing applications in automobile factories, deploying simple tests in our own factories, and AI teams promoting robots to work at home, but since overall efficiency and versatility are still not high enough, they have not been promoted on a large scale. Wang Xingxing believes that moving towards an intelligent ChatGPT moment would ideally take 2 to 3 years, or 5 to 10 years longer. Reaching this point means that in 80% of unfamiliar scenes, robots can successfully complete about 80% of tasks through voice or text commands. Currently, the core challenge is aligning the input and output of a large AI model with a real robot.

2d ago

Yao Shunyu reorganizes Tencent's multi-modal route: closer to Liang Wenfeng and away from Li Feifei

Text | Miao Zheng Editor | Wang Jing Source | Letter AI Tencent's mixed element multi-modal team has undergone another personnel change. According to media reports, Lin Xudong, who was responsible for xAI's multi-modal understanding, has left xAI and joined Tencent's mixed element as the head of the multi-modal content generation algorithm. The reason this personnel news is worth paying attention to is that it takes place in the context of continuous adjustments of mixed and multi-modal teams. Over the past period of time, news of the departure of the person in charge, the transfer of researchers, and the addition of new members came out one after another within the mixed yuan. Hu Han, the former head of multimodal understanding, left his career to start a business, and Tian Yonglong and others joined Tencent. The reporting relationship between the original multi-modal team also changed with the integration of the big language model department and the multimodal model department. However, does this mean that Tencent's multi-modal team is “changing the dynasty” is currently unable to draw a direct conclusion. What can be confirmed by public information is that mixed forces have indeed experienced personnel movements and organizational restructuring. The rumor of Lin Xudong's addition is more like a new signal in this adjustment: Tencent is recombining the two routes of multimodal understanding and content generation. So the question is, what exactly did Lin Xudong come from, and what abilities can he add to Tencent? And is Tencent's multi-modal approach shifting from “generating content” to Yao Shunyu's more biased “understanding context and acting in the world”? What is Lin Xudong's origin and what can he do after joining Tencent? According to public information, Lin Xudong graduated from Tsinghua University in 2018 and then went to Columbia University to study for his doctorate. While studying at the blog, his research interests included embedded learning, video analysis, and generative models. He also participated in the Vx2Text project in collaboration with Columbia University and Facebook AI. V indicates video, x indicates unknown, can be sound, voice, or even ambient sound. 2 represents TO, and Text represents subtitles. Its logic is to first convert different modes such as video and sound into vectors similar to “language tokens”, then uniformly feed the language model for fusion, and finally generate open text by an autoregressive decoder. Transformer can only understand tokens, so AI essentially doesn't understand video and audio file formats, making it even less likely to convert them into text. For example, if a dog jumps into the water next to a swimming pool, Vx2Text's video recognizer (V) will output keywords: dog, jump, pool; sound reader (x) will output: sound of water, fluttering. Although the product function of Vx2Text is “generation,” the core difficulty of the product is “understanding.” Of course, Vx2Text doesn't simply “translate” a screen into a few sentences. Models need to recognize people, objects, movements, and events from videos, understand how these things change over time, and finally organize visual information into language. After graduating from his PhD, Lin Xudong joined DeepMind and participated in Gemini-related multi-modal pre-training and post-training work. In 2025, he also joined xAI. According to public information, it is responsible for the direction of multimodal understanding and participating in the training of multimodal content understanding and generation models. Now that he has joined Tencent Hybrid, he will be responsible for the hybrid multi-modal content generation algorithm. Lin Xudong was added not so much to improve the performance of mixed-element multi-modal generation, but rather to solve a problem that plagues all multimodals — understanding. The previous generation model was more like a picture maker. Give it a hint, and it can generate an image or a video. But as long as users make more complex requests, the model just can't keep up. For example, the characters change in the long video, the shape of the object is not consistent before and after, the camera movement does not match the spatial relationship, etc. It's not because the model doesn't generate, but because it doesn't remember and understand the world steadily. Therefore, putting Lin Xudong in the position of multi-modal content generation is probably because he “translated” multi-modality into something AI can understand. Lin Xudong's addition can only be clearly seen in a larger context. That is, now Tencent's mixed element is reorganizing its multi-modal route. In January 2025, Tencent Outstanding Scientist (Tencent Distinguished Scientist) Hu Han succeeded Liu Wei, who had previously left his job, and was fully responsible for the research and development of mixed-element multi-modal models, and also served as Tencent's mixed-element big model Tech Lead. Tencent's internal organization was adjusted in the second half of 2025. He transferred from the Multimodal Model Department to the “Frontier” Frontier Technology Research Group under the Big Language Model Department. The title was changed to Head of the Multimodal Understanding Direction, and the reporting line was also changed to report to Yao Shunyu. The actual position changed from “the head of an independent department” to a “big language model...

2d ago字母AI#AI #Li Feifei #Liang Wenfeng #Tencent

Byte Seed reorganization: pre-training, RL consolidation, separate groups of office agents

Comparative news, AI news, ByteDance Big Model Team Seed completed a new round of organizational adjustments. Seed Foundation Model established four new first-level departments to re-merge data and post-training teams previously scattered in the fields of text, code, vision, and speech. The core change is to unify pre-training data with reinforcement learning, and at the same time split post-application training into two lines of Work and Chat. Pretrain Data unifies the multi-modal data for Omni models and the data required for pre-training very large models. Horizon RL focuses on strengthening learning and improving basic model capabilities. Product Posttrain-Work is aimed at the office and B-side, and focuses on optimizing the ability to call models, operate computers, and perform long-term tasks. The original Application team changed its name to Product Posttrain-Chat and continues to be responsible for the C-side conversation model. All four departments reported to Wu Yonghui. In the past, Seed was more divided into teams based on text, code, vision, voice, etc., and each had data and post-training personnel. Now Byte wants the next generation model to go directly to Omni, so that different modes enter the same base. Seed was also previously discussing training models with more than 5 trillion parameters. The larger the model and the more modes, the easier it is for the original scattered R&D methods to cause repeated investment. Similar adjustments have also appeared on Tencent. Tencent merged the mixed big language model and multi-modal team in July, with Yao Shunyu under unified responsibility. Byte also set up a special post-work training department this time, indicating that Office Agent has become a product line optimized separately by the basic model team.

3d ago

Li Feifei warns that anti-AI sentiment in the US is heating up: if there is no positive development path, the world will be affected

Comparing the news, AI pioneer Li Feifei said that the tech industry needs to better explain the value brought by artificial intelligence to the public and warned that growing anti-AI sentiment in the US could pose a risk to global AI development. In an interview with Bloomberg, Li Feifei said that technology practitioners need to strengthen communication with the public to show the positive impact AI technology can bring. She pointed out that “if we don't show a positive attitude and a positive direction of development, everyone will be affected”. The US has an important influence in the field of global technology, and if the US fails to show a “positive AI attitude and development path,” it will eventually affect the entire world. As an important researcher in the field of computer vision, Li Feifei has continued to promote AI technology from laboratories to real-world applications in recent years. World Labs, the AI startup she founded, is dedicated to developing spatial intelligence (Spatial Intelligence) technology to explore AI's ability to understand and interact with the real world. Li Feifei believes that the AI industry is currently facing not only technical challenges, but also issues such as social acceptance, public trust, and the regulatory environment. With the rapid spread of generative AI, concerns about employment impacts, data privacy, and security risks continue to increase, and American society's backlash against AI development is expanding. She emphasized that the AI industry needs to be more proactive in explaining how technology can improve fields such as healthcare, scientific research, and productivity, rather than just focusing on technology competition itself. In recent years, the world's major technology companies have continued to increase investment in AI, promoting the development of technologies such as big models, AI infrastructure, and intelligent agents. But at the same time, issues of AI regulation, employment substitution, and social impact have also become the focus of attention of policy makers and the public.

3d ago

Tesla launches large beanbag model

Compared to the news, Tesla launched a large model of bean bags. According to reports, Tesla cars are now being promoted one after another. (Volcano engine)

3d ago