China AI news, from the source
Top stories
Updated Oct 2, 2026 Share this briefingToday's brief is being compiled. Verified Events keep landing in the live update ledger below and in the full Event History.
Browse the Event HistoryMore China AI news
Evidence-qualified product, pricing, policy, capital and deployment Events, ordered by when they happened. Research appears only after an impact review; unverified Signals stay separate.
18 stories · last 30 days
Newest first. Wider windows expand what is available; open Event History for the full period.
Sep 28, 2026 1 story
Manus AI releases Manus 2.0
Manus AI released Manus 2.0, announced on the Manus Blog.
Why it matters The release of a new major version by Manus AI signals continued product development in the AI agent space.
Sep 27, 2026 2 stories
Xiaomi MiMo released MiMo-V2.6-Flash-MOPD on Hugging Face
Xiaomi MiMo released MiMo-V2.6-Flash-MOPD on Hugging Face, a sparse Mixture-of-Experts model with 309B total and 15B activated parameters, 1M-token context, and multimodal support…
Why it matters Developers outside China can now access a 309B-parameter multimodal MoE model under an MIT license, enabling local deployment and fine-tuning without API costs or data-sharing constraints.
Xiaomi MiMo released MiMo-V2.6-Pro-MOPD on Hugging Face
Xiaomi MiMo released MiMo-V2.6-Pro-MOPD on Hugging Face, an upgraded checkpoint of MiMo-V2.6-Pro-RL that mitigates tool-call repetition and extends to domains like game developmen…
Why it matters Developers using Xiaomi MiMo for agentic applications can now access a checkpoint that reduces tool-call repetition, which can lower token costs and improve task completion rates in long-horizon…
Sep 24, 2026 2 stories
Alibaba Cloud announces Qwen 4 in training and first cloud regions in Türkiye, Finland, and the Netherlands
Alibaba Cloud announced at its 2026 Apsara Conference that next-generation Qwen 4 is in training, with Qwen 4.5 and Qwen 5 projected to scale up to 5 to 10 trillion parameters. Th…
Why it matters Alibaba Cloud's expansion into Türkiye, Finland, and the Netherlands will give European and Turkish businesses direct access to Alibaba Cloud services, reducing latency and addressing data…
Alibaba Unveils Agentic Computer, AI Wearables and More at 2026 Apsara Conference
At the 2026 Apsara Conference, Alibaba unveiled an agentic computer, latest AI glasses, earbuds, and an upgraded note-taking device. Qwen Book unifies models, OS and cloud in an a…
Why it matters Alibaba is expanding from cloud and models into consumer AI hardware, competing with other tech giants in the wearables and personal computing space.
Sep 23, 2026 1 story
CAC publishes filing information for three on-device generative AI services
The Cyberspace Administration of China (CAC) published filing information for three generative AI services that provide on-device (mobile phone side) capabilities. The announcemen…
Why it matters Mobile phone manufacturers and AI service providers in China must complete CAC filing before offering on-device generative AI features to users; this announcement confirms that at least three such…
Sep 22, 2026 1 story
Alibaba Launches Qwen Intelligence to Power Next-Generation Agentic Smartphones
Alibaba unveiled Qwen Intelligence, a full-stack agentic solution for smartphone manufacturers. HONOR is the first partner, co-developing domain-specific models. HONOR's Magic9 Se…
Why it matters Alibaba is expanding its AI offerings into consumer devices through partnerships with smartphone makers, starting with HONOR.
Sep 21, 2026 4 stories
Xiaomi MiMo released MiMo-V2.6-Distill-Qwen-9B on Hugging Face
Xiaomi MiMo released MiMo-V2.6-Distill-Qwen-9B, a 9B agentic model fine-tuned from Qwen3.5-9B, on Hugging Face under the MIT license. The model covers coding, general agent tasks,…
Why it matters Developers outside China gain immediate access to a 9B open-weight agentic model under the MIT license, enabling them to build and fine-tune agent applications without licensing costs.
Xiaomi MiMo released MiMo-V2.6-Flash-RL on Hugging Face
Xiaomi MiMo released MiMo-V2.6-Flash-RL, a sparse Mixture-of-Experts model with 309B total and 15B activated parameters, on Hugging Face under the MIT license. It supports 1M-toke…
Why it matters Developers outside China can now use a 309B-parameter open-weights model with 1M context and omnimodal input under the MIT license, reducing the cost of building long-horizon agents compared to…
Xiaomi MiMo released MiMo-V2.6-Pro-RL on Hugging Face
Xiaomi MiMo released MiMo-V2.6-Pro-RL, a 1.02T-parameter sparse MoE model with 42B activated parameters, on Hugging Face under an MIT license. It supports 1M-token context and tex…
Why it matters Developers outside China can now access and modify a frontier-scale open-weights model with 1M-token context and omnimodal capabilities under a permissive MIT license, reducing the cost and…
Alibaba Releases Qwen3.8-Omni-Flash with 1M-Token Context for Audio-Visual Agents
Alibaba unveiled Qwen3.8-Omni-Flash, a native omni-modal model with a 1-million-token context window, targeting music video creation, film production, multimedia commentary, and r…
Why it matters Alibaba is positioning Qwen3.8-Omni-Flash for creative and media production use cases, competing with other multimodal models targeting content generation and real-time interaction.
Sep 20, 2026 2 stories
Zhipu AI releases ZCode open-source coding agent repository
Z.ai (Zhipu) published the ZCode repository on GitHub. The repository contains a coding agent workbench with desktop app, browser interface, and terminal agent. It includes client…
Why it matters Zhipu AI is expanding its developer tooling by open-sourcing a coding agent workbench, competing with other AI coding assistants.
Qwen released Qwen-Image-2.1-PE-I2I on Hugging Face
Qwen released Qwen-Image-2.1-PE-I2I on Hugging Face under the Qwen Research License Agreement. It is a fine-tuned Qwen3.5-VL 9B model that rewrites vague editing instructions into…
Why it matters Qwen continues to expand its open-weights image generation ecosystem by releasing auxiliary models like prompt rewriters alongside the main generation model.
Sep 18, 2026 1 story
Qwen releases RecreationWorld: scalable verifiable environments for hybrid computer-use agents
Qwen (Alibaba) open-sourced RecreationWorld on GitHub under the MIT License. It is a five-platform framework for hybrid computer-use agents that interleave GUI exploration, coding…
Why it matters The benchmark results show a wide performance gap between frontier models, with GPT-6 Astra leading at 58.06 average and Qwen's own Qwen3.7-Plus at 9.15.
Sep 17, 2026 1 story
SenseTime Releases SenseNova U1.5, an 8B-MoT Unified Multimodal Model Without External Vision Encoders or VAEs
SenseTime introduced SenseNova U1.5, a natively unified multimodal model built on an 8B-MoT architecture that understands, reasons about, and generates visual content without exte…
Why it matters SenseTime is positioning itself in the emerging category of natively unified multimodal models, competing with approaches that use external vision encoders or VAEs.
Sep 14, 2026 1 story
China releases AI Safety Governance Framework 3.0
The Cyberspace Administration of China (CAC) published the AI Safety Governance Framework 3.0.
Why it matters The updated framework signals continued regulatory development for AI in China, which may affect compliance requirements for AI developers and deployers.
Sep 11, 2026 1 story
Shanghai AI Laboratory released Atria-Dawn-Preview on Hugging Face
Shanghai AI Laboratory released Atria-Dawn-Preview, a 744B-parameter MoE agentic model built on GLM-5.2, on Hugging Face under an MIT license. It is available for download on Hugg…
Why it matters Developers outside China can now download and modify a 744B-parameter agentic model under a permissive MIT license, reducing the cost and legal friction of building agentic applications compared to…
Sep 10, 2026 1 story
DeepSeek releases DeepSeek-V4.1-Flash with 890-byte per-token KV cache
DeepSeek released DeepSeek-V4.1-Flash on Hugging Face. It is a multimodal Mixture-of-Experts model with 552B backbone parameters, 8B activated parameters per token during prefill …
Why it matters The 890-byte per-token KV cache is a 4-fold reduction from DeepSeek-V4-Flash and 437-fold from DeepSeek-V1, targeting cost efficiency for input-heavy agentic workloads.
Sep 1, 2026 0 stories
Manus Resumes Independent Operations
Manus announced it is resuming independent operations as an independent agent lab, continuing to develop general AI agents.
Why it matters Manus positioning itself as an independent agent lab indicates a strategic move to operate autonomously in the competitive AI agent market.
Aug 31, 2026 0 stories
DeepSeek released DeepSeek-V4-Flash-Vision-Exp, its first experimental multimodal model in the V4 family
DeepSeek released DeepSeek-V4-Flash-Vision-Exp on Hugging Face under the MIT license. It is an image-text-to-text model built on DeepSeek-V4-Flash with added visual modules. Compa…
Why it matters DeepSeek is extending its open-weights model family into multimodal capabilities, directly competing with other open multimodal models.
Aug 27, 2026 0 stories
Tencent open-sources Hy4-preview, a 770B-parameter MoE flagship model
Tencent released Hy4-preview, a 770B-parameter Mixture-of-Experts model with 49B activated parameters and 1M context length, under Apache-2.0 on Hugging Face, ModelScope, GitCode,…
Why it matters Developers outside China gain immediate access to a frontier-scale open-weights model with Apache-2.0 licensing, enabling self-hosted deployment and fine-tuning without API costs or data-sharing…
Aug 25, 2026 0 stories
Zhipu AI released GLM-5.3-Flash, a 320B-parameter multimodal model with 18B active parameters, on Hugging Face
Zhipu AI (via zai-org) released GLM-5.3-Flash on Hugging Face under the MIT license. The model has 320B total parameters and 18B active parameters, is natively multimodal (image-t…
Why it matters The release of GLM-5.3-Flash with 18B active parameters out of 320B total parameters signals a trend toward sparse architectures for cost-efficient inference.
Zhipu AI released GLM-5.3-BF16 open-weights model on Hugging Face
Z.ai (Zhipu) released GLM-5.3-BF16 on Hugging Face. GLM-5.3 uses the same base model as GLM-5.2, with gains from post-training. On Terminal Bench 3.0, GLM-5.3 scores 28.3 vs GLM-5…
Why it matters The release positions GLM-5.3 as a strong open-weights competitor in coding and cyber tasks, with benchmark scores approaching or exceeding some proprietary models like GPT-5.6 Sol on Terminal Bench…
SenseNova U1.5 Lite image model released via SenseTime Token Plan
SenseTime announced that average daily Token usage of the SenseNova Token Plan has surpassed 500 billion, and the next-generation image creation model SenseNova U1.5 Lite is now a…
Why it matters The announcement highlights SenseTime's strategy of using a token-based plan to distribute models to developers, with the 500 billion daily token usage indicating significant developer adoption.
No additional verified events in this window. Try a wider period or open Event History.
View All eventsNew Chinese AI models
All models- XiaomiMiMo-V2.6-ProSep 21, 2026 · Open weights
- DeepSeekDeepSeek-V4.1-FlashSep 10, 2026 · Open weights · Global API
- Zhipu AIGLM-5.3-FlashAug 26, 2026 · Open weights · Global API
- AlibabaQwen3.8-Flash-NextAug 26, 2026 · Open weights · Global API
- Zhipu AIGLM-5.3Aug 18, 2026 · Open weights · Global API
- AlibabaQwen3.8-27BAug 14, 2026 · Open weights · Global API
Explore Chinese AI
- Chinese AI modelsAccess, weights and licenses.
- Chinese AI companiesLabs, robotics and chip makers.
- China AI policyGenerative AI rules, filings, labeling and export controls.
- Attention IndexEnglish and Chinese source attention.
- ExplainersWhat AIGC means, the six AI tigers and China's AI rules.
- China AI LinesOpen weights, the price war, robotics, chips and regulation.
Your reading list
Save stories. Free, across devices.
How we choose and check storiesEvidence gates, source independence, and editorial boundaries
Start from primary facts along model capability, agents, and commercial validation to find decision-moving inflections. Facts, analysis, and outlook stay labelled separately.
Currently tracking 76 sources. Primary sources first · Facts / Analysis / Forecasts layered · Evidence traceable
How confidence is labeled
- Officially confirmedOfficially confirmed: official notices, papers, GitHub, or regulatory filings.
- Cross-checkedCross-checked: at least two independent sources.
- Public reportPublic report: from open media without official material; not counted as verified.
- Live signalLive signal: source observation not yet verified; excluded from verified counts.