Meituan open-sources LongCat-2.0, simultaneously releasing inference code for domestically produced SIM cards.
Related
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.
Sources say OpenAI, Anthropic, and Google are offering hefty computing power subsidies to startups to compete for enterprise clients.
According to a report by Odaily Odaily, citing sources cited by The Wall Street Journal, OpenAI, Anthropic, and Google are offering startups hundreds of thousands of dollars worth of computing resources and other incentives to attract new enterprise customers. (Jinshi)
OpenAI Codex will integrate with the next-generation flagship model GPT-5.6 Sol Ultra.
According to Beating, Thibault Sottiaux, head of core products at OpenAI, confirmed on social media that the Ultra version of the next-generation flagship model, GPT-5.6 Sol, will be integrated into Codex. Previously, some users complained that OpenAI's decision not to include GPT-5.5 Pro in Codex was a major mistake; if GPT-5.6 Ultra were included, developers wouldn't even need to pay for Claude anymore. Sottiaux subsequently confirmed publicly that the Ultra version is indeed in Codex's plans.
MIRA, an open-source 5B multiplayer world model, uses DINOv3 representation to mitigate long-term drift and can simulate 2v2 battles in Rocket League in real time.
According to Beating, AI research firm General Intuition, in collaboration with French AI lab Kyutai and Epic Games, has launched MIRA, a multiplayer interactive world model. As a generative game simulator supporting real-time multiplayer interaction, MIRA can simulate 2v2 battles in Rocket League in real time, based solely on historical footage and player button presses, without requiring a physics engine, rendering engine, or explicit 3D representation for inference. Unlike the "decoupling of logic computation and image rendering" approach adopted by companies like Odyssey, MIRA takes a generative simulation approach based on video latent space. MIRA boasts 5 billion parameters, and its core design builds the latent prediction space on a frozen general-purpose visual encoder, DINOv3-L. Leveraging pre-trained visual features, the generated latent states can more stably fall within the effective representation space, significantly mitigating image drift and divergence during long-term prediction. For multi-screen alignment, MIRA stitches the latent images from four player perspectives into a unified grid, enabling spatial attention mechanisms to operate naturally across viewpoints and improving the spatial consistency of vehicles, the soccer ball, and key events across multiple perspectives. The Action Dropout introduced during training also helps the system complete the game behavior of vehicles not controlled by commands when parts of the motion flow are missing. Currently, MIRA can run in real time at 20 frames per second on a single NVIDIA B200 graphics card. The team has open-sourced the training and inference code and released the Rocket Science dataset, which contains 1,000 hours of matches, approximately 4,000 hours of video, motion flow, and physics data from four perspectives; the complete training of the model used approximately 10,000 hours of clean match data.
Tencent's Hunyuan 3.0 official version open source: Switching to Apache 2.0 removes overseas restrictions, halving the illusion rate.
According to Beating's monitoring, Tencent officially released the final version of its 295B parameter Hybrid Expert Model (MoE) 3.0. The most significant change is the switch to the more permissive Apache 2.0 open-source license, removing the previous regional restrictions that prohibited its use in the EU, UK, and South Korea. When the preview version was released in April, many overseas or multinational teams were deterred from using it due to strict regional restrictions and the 100 million monthly active user limit. The final version removes all compliance obstacles and has undergone significant optimization to address deployment challenges: it incorporates a fast and slow thinking mechanism and adds a 3.8B parameter multi-token prediction (MTP) layer for parallel generation, effectively reducing inference latency. After fine-grained data cleaning and training constraints, the final version reduced the illusion rate of large models from 12.5% to 5.4%, and the error rate in multi-round interactive testing from 17.4% to 7.9%. To address the persistent problem of tool calls easily going astray during agent development, Hunyuan 3.0 has also made targeted improvements. Whether in mainstream scaffolding tools like Cline or CodeBuddy, the accuracy fluctuation of cross-framework tool calls is kept below 4%, resulting in more stable output. The officially released FP8 quantized version also lowers the GPU memory barrier for local deployment and fine-tuning.
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.