China AI

AIAnthropic reports that Zhipu AI's GLM-5.3 can autonomously build end-to-end cyber exploits and is released without meaningful safeguards against misuse.

Why it matters: The report shows how open-weight safeguards fail under simple bypasses, offering concrete test figures for judging misuse risk in released models.

AIZhipu AI says it has deleted all cloud data tied to a ZCode upload controversy, with two third-party organizations verifying the deletion.

AIAlibaba unveiled the Zhenwu V900 AI chip at its 2026 Apsara Conference, claiming three times its predecessor's performance and support for clusters of up to 500,000 cards.

AISparse attention cuts per-operation KV cache reads but does not reduce overall memory capacity, so top-k cache misses still depend on HBM. SemiAnalysis's InferenceX estimates GB200 at about $0.044 per million total tokens at 150 tokens per second, roughly 12% below MI355X running ATOM at $0.049.

AIKling AI says its Kling 4.0 Flash is live now for Ultra Yearly subscribers, with the full Kling 4.0 coming this October.

AIImport AI 474 is a research newsletter by Jack Clark that surveys four developments and one fiction piece.

AICreator CHunye produced the Japanese-style zombie short drama DUHAI entirely on his own from Episode 3 onward using SenseTime's Seko AI video creation agent.

AIBona Film Group says Sanxingdui: Future Past, a 100-minute theatrical feature that uses AI throughout production, will screen nationwide on October 23, 2026.

AIModelScope released LayerExtract and LayerRemove, two Qwen-Image-2.1 LoRAs for editing flattened images.

AIEmber-1, a model built on Kimi K3 through post-training, reportedly reasons 40% more concisely while keeping the same quality and running 40% faster and cheaper.

AIMiniMax has made MiniMax-M3.1 Flash Preview available on its Token Plan, targeting teams with high-volume, latency-sensitive workloads.

AIAnthropic released Claude Opus 5.5 at about 40% lower cost than Opus 5, priced at $4/$20 per MTok input/output.

AIXiaomi has released MiMo-V2.6-Flash-MOPD on Hugging Face, an upgrade of the MiMo-V2.6-Flash-RL checkpoint that fuses several domain-specialized teachers into one model.

AIXiaomi MiMo reports that tool-call repetition in MiMo-V2.6 reached over 0.05% of responses across agent harnesses, causing stalled agents and wasted context.

Why it matters: The post traces an agent failure to a reward blind spot and compares the costs of two fixes, offering a transferable debugging method for RL-trained tool-calling models.

AIXiaomi has released MiMo-V2.6-Pro-MOPD, an upgrade of the MiMo-V2.6-Pro-RL checkpoint that fuses several domain-specialized teachers into one model via MOPD2 and targets tool-call repetition.

AIXiaomi released an open-source RL environments repository, MiMo-V2.6-RL-oss, on Hugging Face.

AIMeituan LongCat has made its LongCat-2.5-Preview model free to try on OpenCode for two weeks. The offer includes a 1M context window, multimodal support, and zero data retention.

AILongCat-2.5-Preview is free on OpenCode for two weeks, offering a 1M context window, multimodal support, and zero data retention. The post does not provide further details on pricing terms or capabilities beyond these listed features.

AIAlexander Doria reports that the released set is a smaller selection of 989 environments for RL training a 9B distilled model, not the full MiMo.

AIShanghai AI Lab's InternLM released Intern-Decision-4B, a multimodal structured decision model fine-tuned from Qwen3.5-4B, which returns answer distributions for multiple questions in one forward pass.

AIShanghai AI Lab's InternLM released Intern-Decision-2B, a multimodal structured decision model fine-tuned from Qwen3.5-2B that returns calibrated answer distributions for multiple questions in one forward pass.

AISemiAnalysis's China Datacenter Model tracks over 1,000 facilities across 60+ operators and puts China's fleet above 24GW, larger than EMEA. The report attributes the buildout to the Eastern Data, Western Compute policy and hyperscale AI demand, which is moving capacity to western hubs such as Inner Mongolia at construction speeds it says the West cannot match.

AIOpenCode says its $60 of usage for DeepSeek v4.1 Flash is now permanent, as part of its "Operation Cheepseek Phase 2" promotion. The post gives no further details on terms, duration, or eligibility.

AIMeituan LongCat is giving every existing user 5 million free tokens to try LongCat-2.5-Preview, with balances viewable on the LongCat platform usage page. Existing token packages work with both LongCat-2.5-Preview and LongCat-2.0.

AIMeituan's LongCat team has released LongCat-2.5-Preview, a natively multimodal model with 1.6T total parameters, about 48B active, and a 1M-token context window.

Why it matters: The announcement lists concrete scale, context, and multimodal specs, plus a named range of agent tasks, which helps readers gauge the preview's scope against other long-context models.

AIThe TileRT and AMD teams reached 469 tok/s single-user decode for GLM-5.3 on 8× MI355X using vLLM. The setup disaggregates work, with vLLM handling prefill and TileRT handling latency-critical decode through vLLM's V1 connector interface.

AIHuawei will likely remain about four years behind Nvidia in AI chip performance and production through 2030, Epoch AI estimates.

AIBaidu's Apollo Go RT6 vehicles have been testing on London's complex streets and traffic patterns for more than a month with Lyft. The post shares video footage of the ongoing testing without disclosing performance metrics or launch timing.

AIModelScope released Qwen-Image-2.1-Fun-Controlnet-Union, a single checkpoint adding eight structural controls, including Canny, Depth, Pose, and Scribble, plus inpainting to Qwen-Image 2.1.

AIinclusionAI has published public training-content summaries on Hugging Face for its Ling and Ring model versions, including Ling-2.0, Ling-2.5, Ling-2.6-1T, Ling-3.0, Ring-2.0, Ring-2.5-1T, and Ring-2.6-1T. The documents, organized under the template associated with Article 53(1)(d) of Regulation (EU) 2024/1689, contain documentation only, not model weights or training datasets.

AITencent Hy Translation has launched, powered by Hy-MT2, supporting 33 languages and 5 Chinese minority languages and dialects.

AINunchux runs MiniMax-H3 on AMD MI355X GPUs, generating 5 seconds of video in 1.3 seconds with up to 26.7x faster inference than SGLang on 8 GPUs.

AITencent Hunyuan extends classical critical-batch-size theory to online LLM reinforcement learning, where models generate their own training data.

AIAnt Ling's Ming-Image-0.1-Design, a text-to-image model built for graphic-design output and legible text rendering, is now available via Novita on OpenRouter. Novita launched it as a Day-0 partner with a free 14-day offer.

AIInferact says vLLM maintainers used megakernel optimization to reach 700 tokens per second per user on TPUv7 running Kimi K3.

AIInferact says its first TPU megakernel for Kimi K3 reaches 709 tokens/s on low-concurrency decode with DSpark speculative decoding, versus 450 tokens/s for its GB200 baseline.

AIUnsloth has passed 500 million model downloads on Hugging Face. Its Qwen3.8-27B GGUF is the most-downloaded model in Unsloth's history.

AIAlibaba's Qwen launched Qwen Intelligence with three mobile agents: a Mobile Planner Agent, a Mobile-Use Agent, and a Mobile Creative Agent.

Why it matters: The post names three mobile agents and their benchmark results, while also releasing the benchmark suite, so readers can check the claims against the reported figures.

AIAlibaba's Qwen team released Qwen-Audio-3.1, upgrading its ASR, TTS and Realtime models and adding TTS-Next and ASR-Next. The post says TTS prices fell about 70%, Realtime about 85%, and ASR up to 95%. More APIs are coming soon.

Why it matters: The post lists the new audio models and price changes together, which helps developers compare the upgraded lineup with their current speech workflows.

AIXiaomi has released MiMo-V2.6 as an open model family under the MIT License, designed for large-scale reinforcement learning.

Why it matters: The post links benchmark results, parameter scale, and a multi-agent RL training run, giving readers concrete figures to compare against other open models.