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.
Related
LingBot-VLA 2.0, the ant-like Lingbo embodiment base model, is open source.
On July 8th, Antminer Technology announced the upgrade and open-sourcing of its next-generation embodied platform model, LingBot-VLA 2.0. As a comprehensive upgrade to the open-source version LingBot-VLA 1.0 released in January of this year, LingBot-VLA 2.0 incorporates 60,000 hours of high-quality real-world physical data during the pre-training phase, covering 20 robot configurations from 17 mainstream robot brands, and expanding support for degrees of freedom such as the head, waist, end effector, and mobile chassis. This results in significant improvements in configuration generalization, degree of freedom support, and deployment efficiency.
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.
Opinion: Open source models account for only 10% of enterprise large-scale model spending, but mature production environments will be dominated by open source models.
According to Beating's monitoring, while public opinion often touts that open-source large models are dominating everything, enterprise spending data presents the opposite picture. Jesse Zhang, co-founder and CEO of Decagon, an enterprise-level AI customer service platform, points out that the share of open-source models in total enterprise spending has now dropped to 11%. This decline stems from the fact that most enterprises' AI applications are still in the early, undefined exploratory stage, thus defaulting to reliance on closed-source models. However, he emphasizes that once application scenarios mature, open-source models will take over production environments with their advantages of extremely low latency and deep fine-tuning. In Decagon's own production environment, 90% of calls have already switched to open-source weighted models. The core driver of this transformation is interaction speed and customization capabilities, not cost savings. In customer service scenarios, a single conversation that takes 8 seconds to finish will completely destroy the product experience. Since leading closed-source labs do not allow fine-tuning of flagship models, and small closed-source models cannot be deeply customized, small-sized open-source models, through fine-tuning for specific tasks, have become the only option to support high-frequency real-time interactions. The future of enterprise AI will see a division of labor: leading closed-source labs will continue to dominate the exploration and discovery of new fields, while open-source weighted models will increasingly take over the actual production of mature businesses. Because model fine-tuning requires extremely high levels of data and talent, the migration from closed-source to open-source will be a slow process lasting several years, during which both will experience sustained growth.
JPMorgan: Open source weight commercialization exhibits a "winner-takes-all" phenomenon; Zhipu target price raised to HK$2,000, MiniMax target price cut to HK$300.
According to BlockBeats, on July 8th, JPMorgan Chase released a research report stating that currently competitive models in the market can expand adoption through open-source weighting and continue to monetize through official APIs, partner channels, enterprise deployments, and workflow products; while weaker models face faster price comparisons and traffic fragmentation. JPMorgan Chase raised its revenue forecasts for Zhipu from 2026 to 2030 by 3% to 9%, and narrowed its adjusted loss forecasts for 2026 and 2027 to RMB 3.711 billion and RMB 3.141 billion respectively. The 2028 forecast was revised from a loss of RMB 1.287 billion to a profit of RMB 2.367 billion. The target price was raised from HKD 1800 to HKD 2000, maintaining an "Overweight" rating. The report believes that the performance of GLM-5.5/6, KimiK3, and DeepSeekV4.1 will be key indicators of whether Zhipu can maintain its leading position. JPMorgan lowered its revenue forecasts for MINIMAX-W (2027-2030) by 2% to 8%, and reduced its target price from HK$400 to HK$300, while maintaining a "neutral" rating. The company noted that the M3 model offers a permanent 50% discount, reflecting that the model has not yet created a significant capability premium for leading domestic competitors. JPMorgan believes that if MiniMax can narrow the capability gap, normalize the discount, maintain API usage, and demonstrate stronger workflow stickiness through MiniMaxCode, its outlook could turn positive.
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.
Meituan open-sources LongCat-2.0, simultaneously releasing inference code for domestically produced SIM cards.
Mars Finance reported on July 6th that Meituan officially open-sourced its trillion-parameter large-scale computing model, LongCat-2.0. LongCat-2.0 has a total of 1.6T parameters and an average activation of approximately 48B. It features deep collaborative optimization in model architecture, chip adaptation, and deployment strategies, specifically targeting domestically produced computing chips with limited GPU memory and bandwidth. (Wide Angle Observation)