According to Mysteel, citing the Meituan Tech WeChat account, Meituan open-sourced LongCat-2.5 on 10 11, 2026, a natively multimodal model for long-horizon agentic tasks, under the MIT license, allowing commercial inference deployment and further training without additional authorization. The model has 1.6 trillion total parameters, activates approximately 480 billion parameters per token, supports one million-token context, and has been optimized in architecture and deployment for domestic compute chips constrained in memory and bandwidth.
According to hot-topic coverage published by YeYuLingFeng Learning Network on 2026 11, 6, Tencent Hunyuan officially released the 11 large model on 4 Hy4, introducing for the first time visual understanding and computer operation (Computer Use) capabilities compared with the 8 Preview version. In a blind test of 213 tasks by 579 financial experts, Hy4 scored a composite of 3.04, higher than several open-source competitors; the hallucination rate fell from 12.95% to 7.2%. The company also acknowledged a significant gap versus top closed-source models, with the weight release date not yet announced.
Ant Lingbo Technology open-sourced the tactile foundation model LingBot‑Touch on 2026/11/10. Pretrained on nearly 5000 hours of real-world tactile data, it is compatible with magnetic tactile, visuotactile, and piezoresistive sensors, unifying heterogeneous signals into tactile Tokens. In real-robot testing, integration improved task success rates by about 14 to 26 percentage points across four action models, and the lite version has about 5300K parameters. The model has been open-sourced on Hugging Face, GitHub, and the ModelScope community.
On 11 9, Zhi Ping Fang (AI Robotics) announced that it had signed a Memorandum of Understanding with the Danantara Development Management Fund (DDMF), a development investment platform under Indonesia's national strategic investment management body Danantara Indonesia, covering cooperation on pilot applications, industrial collaboration, and localized deployment of general-purpose intelligent robots and embodied AI technology in Indonesia. The source states that Zhi Ping Fang is the first and currently the only robot enterprise on the list, with manufacturing, logistics, and public services launching first and gradually expanding into multiple strategic priority industries in Indonesia.
According to Yicai, OpenAI announced on local time 11 4 the acquisition of developer-tools startup Pierre Computer Company, whose Code Storage, Trees, and Diffs tools have been integrated into its coding agent Codex; the deal amount was not disclosed. On the same day, Cat Wu (Catherine Wu), product lead for Anthropic's Claude Code and Cowork, left to join OpenAI. Combined with the simultaneously announced proprietary Git backend Pando, OpenAI is moving down the stack into development infrastructure such as code storage and version control, though it has not announced a direct rivalry with GitHub.
On 11 2, the 2026 World Internet Conference Leading Science and Technology Award was presented in Wuzhen, and Huawei's Ascend 950 SuperNode project won. Based on the UnifiedBus interconnect protocol, each node is configured with 1024 NPU cards, over 1300 UnifiedBus interconnect chips, and over 4000 optical modules. Officially, the 1024-card SuperNode's training compute utilization can reach 45%, with inference latency as low as 5 ms. Over 1000 Ascend SuperNode systems have been deployed, serving internet, telecom, finance and other industries.
According to a People's Finance News report on 11月5, Huawei Cloud's MaaS platform will update and launch the openPangu-2.1-Pro flagship model on 11月15, with a focus on improving ArkTS code generation quality in HarmonyOS native app development; core products including the agent platform AgentArts, CodeArts, and OfficeAce will simultaneously integrate the model. The report also notes that earlier claims that Huawei Cloud had completely discontinued sales of the Pangu large model were a misreading.
At the 9th China International Import Expo, which opened on 11 5, Zhiyuan Robotics showcased the full-stack technology achievements released at its 2026 Innovation Day and introduced the Worken (valid intelligent work hours) evaluation concept, measuring robots' real output across three dimensions: active device scale, autonomous working duration, and task completion efficiency. The expo publicly revealed two new scenarios for the first time—smart guidance at Peking University People's Hospital and hotel interactive services. As of 2026 10, cumulative valid intelligent work hours exceeded 6.20,000, spanning five major directions: culture, tourism and entertainment; urban transportation; commercial retail; industrial manufacturing; and public livelihood services.
According to Beijing Business Today, on 11 6, Zhiyuan released the updated version of its bipedal humanoid robot Lingxi X2, in a Youth Refreshed Edition and a Flagship Refreshed Edition, with total DOF of 30 and 31, reaching 47 with the optional 16-DOF dexterous hands. The Flagship Edition carries an NVIDIA Jetson Orin 16GB module (compute power 157 TOPS), a 3D LiDAR and a depth camera, with a 18-month warranty; the Youth Edition has baseline compute power with only an interactive camera, and a 8-month warranty. The new product targets exhibition venue tours, in-store services, multi-robot commercial performances, research education, and other scenarios.
At CoRL 2026 held in Austin, Texas, USA, Fourier exhibited its third-generation humanoid robot GR-3 and, for the first time, demonstrated a Physical AI data collection kit jointly developed with Yihang Lingguang and Qineng Robotics. Built on the GR-3 platform, the kit integrates a six-camera vision head ring and a tactile + EMF data glove, capable of simultaneously capturing first-person vision, motion posture, and tactile data, and is open to the global research and industrial ecosystem.
On 11 9, Google released TPU 8tx and TPU 8i Max, targeting confidential inference and ultra-long context processing respectively. TPU 8tx reuses TPU 8t chips and integrates hardware-level confidential computing, with inference latency reduced by up to 50%; TPU 8i Max integrates two TPU 8i chips in a single package, doubling compute, SRAM, and HBM bandwidth. Google also expanded its software stack with TorchTPU, the agent toolset, Agent Compute Engine, TPU Sync, and Cloud Pathways. Four eighth-generation TPU systems have now been released this year.
The Qwen4 release window relayed by outside observers from 2026 11 2 to 8 has passed, and Alibaba's official written statement remains only the 9 22 remark that it is "in training," with no release date, model specs, pricing, or license announced. Verifiable facts in the article include the four model names from the 9 22 live screen projection, the context change merged in the Qwen CLI repository on 10 28, and the official poster data from 11 6. The article recommends moving Qwen4 off the release schedule and keeping only a watch position.