Tencent's Hunyuan 3.0 official version open source: Switching to Apache 2.0 removes overseas restrictions, halving the illusion rate.
Related
Former OpenAI Chinese researcher Tian Yonglong joins Tencent Hunyuan team
According to Beating's monitoring, Tian Yonglong, a former member of the OpenAI technical team, has confirmed joining Tencent's Hunyuan team and will participate in the research and development of visual language models. Tian Yonglong graduated from Tsinghua University with a bachelor's degree and received his Ph.D. from MIT. He previously served as a senior research scientist at Google Research and Google DeepMind, primarily researching computer vision and generative models. This is another top-tier Chinese AI researcher recruited by Tencent Hunyuan, following the recruitment of Chief AI Scientist Yao Shunyu last December. Tencent recently launched the official version of the Hunyuan Hy3 model, led by Yao Shunyu, under an open-source license. With Tian Yonglong's addition, Tencent Hunyuan's talent pool in multimodal and visual model development will be further strengthened.
Tencent's "Toast" iOS version officially launched
According to Mars Finance, on July 6th, Tencent's AI application generation and inspiration co-creation platform "Toast" officially launched its iOS version. As a Vibe Coding product with zero barriers to entry, users only need to describe their needs in natural language to generate and customize their own apps with Toast. It is understood that Toast will fully open its one-stop listing service, commercial component capabilities, and more API capabilities to the developer ecosystem in August.
Tencent shares surged over 5% intraday, marking a narrative reversal: WorkBuddy users increased by two orders of magnitude, and WeChat's AI "Xiaowei" opens up greater possibilities.
According to BlockBeats, on July 6th, Bitget data showed that Tencent Holdings' Hong Kong-listed shares rose over 5% intraday, reaching HK$453. Tencent, once criticized for its "slow AI development," is experiencing a narrative reversal. In May, at the shareholders' meeting, Ma Huateng frankly admitted, "A year ago we thought we were on a ship, but then we found it was leaking, and now we're standing on it but can't sit down." However, in just a few months, WorkBuddy's monthly active users (MAU) exceeded 20 million, daily active users (DAU) exceeded 13 million, and the DAU/MAU ratio reached 65%-75%, with stickiness comparable to Slack-level office tools. With 43 versions released in three months and updates continuing even during the May Day holiday, Tencent internally positions it as its third phenomenal product after QQ and WeChat. WorkBuddy's success hinges on two key barriers: first, its extremely simple installation—a desktop client plus a WeChat mini-program, with commands sent from the phone after WeChat authorization and executed automatically on the PC, completely bypassing the high command-line deployment hurdles of similar foreign products; second, the activation of Tencent's entire ecosystem in the Agent era—integrating with Tencent Docs, Tencent Meeting, WeChat Work, and over 30 external tools, expanding the SkillHub skill library from 79,000 to 790,000, allowing for automatic data retrieval, file generation, and distribution with a simple command like "create a PPT of last week's sales data and send it to colleagues at a meeting." This leap from a chat bot to a tool that gets things done transforms a fragmented toolchain into a competitive advantage. Furthermore, this groundbreaking product, the first in Tencent's history to conduct on-the-ground promotion, reflects a genuine commitment to penetrating the B2B market. Even greater potential lies in WeChat AI. WeChat's AI assistant, "Xiaowei," launched a small-scale gray-scale test on June 20th, supporting voice or text messaging, Moments posting, and accessing mini-programs for appointment booking and food delivery. It even generates mini-program prototypes using natural language. This marks the first time this super app with 1.43 billion monthly active users has entered the intelligent agent arena. JPMorgan Chase estimates, based on probability weighting, that WeChat's AI Agent could generate approximately 126 billion yuan in incremental revenue and 88 billion yuan in incremental operating profit by 2030. However, Goldman Sachs warns of three major concerns: the potential resource duplication from the parallel operation of its self-developed WeLM model and the mixed-source model; the possibility that inference costs after full rollout could erode 5%-17% of the projected adjusted operating profit for Q4; and the unclear short-term monetization path. The core difference between the two institutions lies in the timing: JPMorgan Chase believes the public beta has transformed WeChat AI from an "abstract option" into a "traceable milestone," while Goldman Sachs believes the short-term cost-benefit asymmetry requires further quarterly verification. In terms of the underlying model, Hunyuan 3 Preview has been integrated into 131 Tencent products, with token usage increasing tenfold compared to the second generation. It ranks first in usage on the OpenRouter platform. However, the prediction that Ma Huateng (Pony Ma) might not be able to maintain his position even after taking the lead still holds true. The ecosystem is the natural foundation for Tencent to act as an agent, but WeChat AI faces complex challenges in handling transactions, advertising, payments, merchants, fairness, and regulation.
Tencent Cloud: DeepSeek-V4 official version "direct from the manufacturer" model is scheduled to be launched in mid-July.
PANews reported on July 3 that, according to a Tencent Cloud announcement, Tencent Cloud's large model service platform TokenHub and intelligent agent development platform plan to officially launch the DeepSeek-V4 official version "direct from the manufacturer" model in mid-July, and simultaneously introduce a peak-valley billing mechanism. DeepSeek-V4-Pro will be charged based on millions of tokens during normal times: 0.025 yuan for cache hit, 3 yuan for inference input, and 6 yuan for inference output; during peak hours, the charges will be 0.05 yuan, 6 yuan, and 12 yuan respectively. DeepSeek-V4-Flash will be charged during normal times: 0.02 yuan for cache hit, 1 yuan for inference input, and 2 yuan for inference output; during peak hours, the charges will be 0.04 yuan, 2 yuan, and 4 yuan respectively.
Meituan open-sourced its trillion-parameter large-scale model LongCat-2.0, and simultaneously released the inference code for domestically developed Chinese card processors.
According to Beating's monitoring, Meituan has officially open-sourced its trillion-parameter large-scale model, LongCat-2.0, with a total of 1.6T parameters and an average activation of approximately 48B, designed specifically for real-world agentic coding tasks. Architecturally, it innovatively introduces LongCat sparse attention and N-gram embedding. The former reduces fragmented memory access through flow-aware indexing and hierarchical indexing, accelerating training and inference with millions of contexts; the latter, while achieving nearly 97% sparsity in MoE, invests 135B parameters into the embedding layer, balancing parameter gains and structural stability. Post-training employs multi-teacher online distillation, categorizing experts into Agent, Inference, and Interaction types, seamlessly integrating them on a domestic computing power cluster through the MOPD architecture. As the industry's first trillion-parameter model to complete inference on a 50,000-card domestic computing power cluster, LongCat-2.0 validates the mature capability of domestic chips to handle complex large-scale model tasks. To address the multiple limitations of domestically produced Chinese chips in terms of memory, bandwidth, and interconnects, Meituan has made breakthroughs in three areas: model, chip adaptation, and deployment. At the model level, ScMoE leverages the core control capabilities of domestically produced chips to achieve physical core-level parallelism for Dense and MoE branches, combined with KV-cache partitioning to alleviate the pressure on ultra-long context memory. At the chip adaptation level, Super Kernel reduces operator startup overhead, and Weight Prefetch hides I/O latency, maximizing hardware utilization under constrained conditions. At the deployment level, PD separation is adopted to balance TTFT and TPOT, along with asynchronous Expert-Parallel load balancing to solve load unevenness under high EP (efficiency level). This open-source release simultaneously provides multiple precision versions, including BF16, FP8, and INT8, and fully opens up inference results optimized for domestic computing power, aiming to enable existing domestically produced cards and even older cards to smoothly deploy trillion-model inference services. --------------------------------- Click the original link below to join the Beating · Lark AI news channel and monitor global AI hot topics and news 24/7.
Theorem proofs are also starting to come down to cost: Mistral open-sourced Leanstral 1.5, costing about $4 per problem.
According to Beating's monitoring, Mistral AI has open-sourced Leanstral 1.5, a model for Lean 4 formal proofs. The model has a total of 119 billion parameters, with approximately 6.5 billion activation parameters. It is licensed under the Apache 2.0 license and offers a free API. Official evaluations show that Leanstral 1.5 solved 587 out of 672 problems on PutnamBench; it achieved 87% and 34% on the abstract algebra benchmarks FATE-H and FATE-X respectively, setting new best performance records for similar models. The average solution cost of Leanstral 1.5 on PutnamBench is approximately $4, lower than the tens to hundreds of dollars cost of some previous systems. As the token budget per problem increases, the number of problems it solves continues to increase; in the AVL tree complexity proof, the model completed the proof after more than 2.7 million token inferences and 22 instances of context compression. In addition to mathematical proofs, Leanstral 1.5 was also used for code verification. The team discovered 11 real bugs in 57 open-source Rust repositories, 5 of which had not been previously reported.