China AI

AIDeveloper Maurice Heumann says his team used over 500 billion tokens and three months of AI agent work to decompile the 2009 game Call of Duty: Modern Warfare 2 into C++. The team checked each rebuilt function by compiling it and comparing its bytes against the original, reaching 99% function coverage and 83% exact matches.

AIDeepSeek founder Liang Wenfeng is considering entering embodied AI, either by building a team in-house or through investments, according to people familiar with the matter.

AIChina Mobile has tripled the processing capacity of its largest intelligent computing hub in Hohhot, Inner Mongolia, to 20,100 petaflops, up from 6,700 petaflops, according to Inner Mongolia Daily.

AIJerry Liu shares a post saying Tencent has open-sourced TencentDB Agent Memory, a memory system for AI agents that runs locally on SQLite without external vector databases or cloud APIs.

AIThe 2026 International Embodied AI Skills Competition will be held in Shanghai on October 23, according to the source. The event is described as a live-robot training and practical showcase platform for embodied intelligence.

AIAnthropic says its new defenses have cut the reasoning tokens distillers can extract from Claude by a factor of 10 to 100 on its own platform.

Why it matters: The post names specific distillation methods and their measured effects on reasoning extraction, which helps readers judge how AI labs protect models from copying.

AIDesign Arena says Mistral Large 4 is the top non-Chinese open-weights model across its Website, Data Visualization, UI Components, 3D Design, SVG, ASCII Art, and Game Dev arenas.

AIAlibaba's Qwen team introduced QwenGyre, an online reinforcement learning framework that trains coding agents through unmodified black-box harnesses on runs of hours and about 700K tokens.

AINathan Lambert says a recent report is the best he has read on China's LLM data economy and how it compares with the US. He says Tencent, Alibaba, and ByteDance, China's leading spenders, do not mirror OpenAI and Anthropic, the leaders in the US market.

AIArena says the top five AI agents now have 94% of delegated tasks confirmed as done by users, up from 80% in July.

AIArtificial Analysis says GPT-6 Luna (max), GPT-6 Astra (max), GPT-6.1 Sol (max), and Claude Opus 5.5 (max with fallback) sit on the Pareto frontier for score versus time per task.

AIOpenBMB announces MiniCPM-V 4.7, a new open-source multimodal model series with a 35B-A3B model and a 1B model.

Why it matters: The post gives concrete on-device speed, memory, and benchmark figures for two new open-source models, useful for judging local multimodal deployment options.

AIOpenBMB introduces Lookahead Expert Scheduling, or LES, in MiniCPM-V 4.7 to speed up running a 35B MoE model on memory-limited devices, where expert weights stored on SSD can stall generation during on-demand loading.

AIOpenBMB announces MiniCPM-V 4.7, which it says scales from mobile devices to PCs and edge systems.

AIOpenBMB's MiniCPM-V 4.7 uses Lookahead Expert Scheduling (LES) to speed up running a 35B MoE model on memory-limited devices with expert weights stored on SSD. LES predicts which experts the next layer may need, computes already-loaded experts while loading the rest in parallel, and needs no additional training.

AIOpenBMB announces MiniCPM-V 4.7, an open-source model it says can scale from mobile devices to PCs and edge systems.

AIZ.ai introduces GLM-Legal, built on GLM-5.3-Flash, which it says matches frontier models on mainstream legal benchmarks.

AIHuawei says its Ascend AI chips have overtaken Nvidia in China without disclosing market share.

AIOpenBMB releases Trident-Qwen3VL-2B, a multimodal embedding model that maps text, images, and fused text-image inputs into one shared space for mixed-modality retrieval.

AIUnitree Robotics CEO Wang Xingxing says OpenAI's recently released GPT-6 has raised expectations that a breakthrough in robotics could come within one or two years.

AIQwen introduces QwenGyre, an elastic reinforcement learning system for extreme-long-horizon agent training, with an Elastic Scheduler that reallocates GPUs between rollout and training.

AIModelScope says DMAD is now open source and enables 1 to 4 step generation for image, video, and joint audio-video models.

AIElorian and Scale AI release Humanity's Sixth Sense (HSS), a 522-question benchmark for intuitive visual reasoning, where humans score 93.1 pass@1 and the best model, GPT 6 Astra, scores 53.6.

AISimon Meyer, creative director and co-founder of PROMPTR Studios, used early access to Kling 4.0 to create an AI-generated music video that went viral.

AIMicrosoft has released Microsoft-Decision-1, a decision model post-trained from Alibaba's Qwen3.5-9B. Microsoft says it outperforms LLMs and the Jev 1.13.0 model on latency and accuracy for structured decision tasks.

AITencent's WorkBuddy workplace platform has integrated Stepfun's Step 5 Preview, available free for two weeks exclusively on WorkBuddy.

AIAnthropic's July report says Claude banned 11.4 million accounts in the first half of 2026, according to a report that circulated again this week.

AIBaidu says Baidu World 2026 will take place November 12 at the Beijing National Convention Center Phase II under the theme "The next act, intelligent organizations," with ticket sales now open on the conference website.

AIThe Semiconductor Industry Association says China accounts for about 5 per cent of equipment and materials directly supplied to the US chip industry when measured by country of manufacture.

AIFormer Apple CEO Tim Cook met China's Guangdong governor Meng Fanli, while COO Sabih Khan visited Foxconn's Zhengzhou factory before the iPhone Duo's October 18 launch.

AIHuawei Noah's Ark Lab and UCL report that Memento 3, built on Claude Opus 5 with frozen weights, cleared all 25 public ARC-AGI-3 games and 183 levels.

AIResearchers from Huawei Noah's Ark Lab and University College London released Memento 3, an agent that cleared all 25 public games in ARC-AGI-3 using 7,518 actions, about 44% of the human total.

Why it matters: The source separates the agent design from the base model and states the public-set limits, so readers can judge what the benchmark result shows.

AIVisincept's WorldIncept won Track 1 video prediction with 72.33 points, edging Tongji's BWM-Turbo (71.49) by 0.84 points across 81 models.

AIStepFun says its Step 5 Preview model is available free in Tencent's WorkBuddy AI office agent from October 12 to October 25, an arrangement it describes as exclusive.

AIBaidu will hold its annual Baidu World conference on November 12 at the China National Convention Center Phase II in Beijing, and ticket sales have opened on the event's official website.

AIA non-peer-reviewed paper led by ByteDance researchers with Princeton, Stanford and UC Berkeley affiliates says DeepSeek's memory-reduction method for long inputs may make models miss information in certain document positions.

AIModelScope announces that the Step 5 Preview, a flagship foundation model built for real-world agentic tasks, is going open source. The post gives no release date and links to a countdown page for the model on ModelScope.

AIZhipu's GLM-5.3-Highspeed model is reportedly available to random Zcode 3.15.1 subscribers, with one user reporting a speed of 490 tokens/s.

AIBlack Lake CEO Zhou Yuxiang says many Chinese factories are skipping full traditional digitalisation and moving directly to AI agents, in the same way China skipped credit cards for mobile payments.

AIA T. Rowe Price survey of 1,004 Hong Kong adults aged 30 and above found 55% expect AI to make their retirement path less predictable than previous generations' paths.