Thesis
Chinese LLM release velocity accelerated sharply in 2024–2026, with at least one benchmark-setting open-source release per quarter from December 2024 onwards. DeepSeek-V3 (MIT, December 2024), DeepSeek-R1 (MIT, January 2025), Kimi k1.5 (February 2025), and Qwen3 (Apache 2.0, April 2026) each set new open-source benchmarks within months of each other. Lab official channels — X accounts and GitHub/HuggingFace pages — are consistently 4–6 hours faster than English media in publishing the technical details that matter for build decisions (benchmark numbers, architecture, license). The key setup for any builder tracking Chinese AI is: X notifications from lab accounts, HuggingFace trending watch, and a GitHub release watch for QwenLM and deepseek-ai. The RadarAI weekly digest then provides benchmark context and licensing verification within the week.
Decision in 20 seconds
| Lab to track | Fastest English release surface |
|---|---|
| Alibaba Qwen | @qwen_lm on X → QwenLM GitHub (github.com/QwenLM) → qwenlm.github.io blog |
| DeepSeek | @deepseek_ai on X → huggingface.co/deepseek-ai → github.com/deepseek-ai |
| Moonshot / Kimi | @MoonshotAI_Kimi on X → moonshot.cn blog → platform.moonshot.cn API docs |
| Zhipu AI / GLM | @zhipuai_cn on X → huggingface.co/THUDM → open.bigmodel.cn changelog |
| MiniMax | @minimaxai_global on X → minimax.io blog → huggingface.co/MiniMaxAI |
Chinese LLM release timeline (2024–2026, reverse chronological)
| Model | Lab | Release date | Key specs | License | Benchmark headline |
|---|---|---|---|---|---|
| Qwen3 series (8 models) | Alibaba | April 2026 | 0.6B–235B; Qwen3-30B-A3B (MoE, 3B active); Qwen3-235B-A22B (MoE) | Apache 2.0 | Qwen3-30B-A3B matches GPT-4o on MMLU/AIME; best open-source cost-efficiency at release |
| DeepSeek-Prover-V2 | DeepSeek | May 2025 | 671B MoE; math-specialized; 7B distilled version released simultaneously | MIT | State-of-the-art on MATH-500 and AMC/AIME formal proof tasks |
| Kimi k1.5 | Moonshot | February 2025 | Long-chain-of-thought; 128K context; multimodal | Proprietary (API only) | Competitive with GPT-4o on long-context benchmarks; matches o1 on AIME (reported) |
| DeepSeek-R1 | DeepSeek | January 2025 | 671B MoE; reasoning-specialized; 1.5B–70B distilled variants | MIT | Matches OpenAI o1 on AIME 2024, MATH-500, and Codeforces; first open-source o1-tier model |
| MiniMax-Text-01 | MiniMax | January 2025 | 456B MoE; 4M token context window (largest at release) | Apache 2.0 (weights); API available | First model with 4M context window; strong on long-document retrieval tasks |
| DeepSeek-V3 | DeepSeek | December 2024 | 671B MoE (37B active); trained on 14.8T tokens; $5.576M training cost (reported) | MIT | Exceeded GPT-4o on SWE-bench Verified and AIME 2024; best open-source model at release |
| GLM-4-Plus | Zhipu AI | September 2024 | 128K context; CogVideoX (video gen) integration; updated instruction following | Proprietary (API); GLM-4-9B open weights | Strong on Chinese instruction-following benchmarks; CogVideoX-5B open-sourced separately |
| Qwen2.5 series | Alibaba | September 2024 | 0.5B–72B; Qwen2.5-72B, Qwen2.5-Coder-32B, Qwen2.5-Math-72B | Apache 2.0 | Qwen2.5-72B top open-source general model at release; Coder-32B matched GPT-4o on coding benchmarks |
| DeepSeek-V2.5 | DeepSeek | September 2024 | Merged DeepSeek-V2-Chat and DeepSeek-Coder-V2-Instruct into unified model | MIT | Improved coding and reasoning over V2; lower cost per token than contemporaries |
| Qwen2 series | Alibaba | June 2024 | 0.5B–72B; 128K context; strong multilingual (29 languages) | Apache 2.0 (most models) | Qwen2-72B beat Llama 3 70B on MMLU, HumanEval, GSM8K at release |
Lab release tracking channels — speed comparison
| Lab | GitHub | HuggingFace | X account | Official blog | English media lag |
|---|---|---|---|---|---|
| Alibaba Qwen | github.com/QwenLM — releases.atom available | huggingface.co/Qwen — model cards go live same-day | @qwen_lm — posts in English; often first channel | qwenlm.github.io — English blog with benchmark context | 3–6h for headline; 6–24h for technical depth |
| DeepSeek | github.com/deepseek-ai — new repos created per release | huggingface.co/deepseek-ai — weights and tech report | @deepseek_ai — English posts; simultaneous with GitHub | deepseek.com/news — English summaries | 4–6h for headline; 12–48h for detailed coverage |
| Moonshot / Kimi | github.com/MoonshotAI — less active than Qwen/DeepSeek | huggingface.co/moonshotai — partial weight releases | @MoonshotAI_Kimi — English posts for major releases | moonshot.cn/news — mixed Chinese/English | 4–8h; less systematic English media coverage than Qwen/DeepSeek |
| Zhipu AI / GLM | github.com/THUDM — GLM-4 open weights | huggingface.co/THUDM — CogVideoX, GLM-4 weights | @zhipuai_cn — primarily Chinese; some English on major releases | zhipuai.cn/news — primarily Chinese | 6–12h; English coverage less consistent; TechNode most reliable |
| MiniMax | github.com/MiniMaxAI — Text-01 weights released | huggingface.co/MiniMaxAI — Text-01, video models | @minimaxai_global — English posts; active since 2025 | minimax.io/news — English blog | 4–8h; growing English media coverage since Hailuo AI international launch |
| Baidu ERNIE | github.com/PaddlePaddle — PaddleNLP; ERNIE not fully open-sourced | huggingface.co/baidu — limited; API-focused | @BaiduResearch — irregular; primarily English for academic work | research.baidu.com — English; yiyan.baidu.com — Chinese | 12–24h; less builder-relevant than Qwen/DeepSeek; Reuters covers major Baidu announcements |
Setting up real-time alerts: step-by-step
Minimum viable alert setup for Chinese LLM releases (takes under 10 minutes to configure):
- X/Twitter: Follow @qwen_lm, @deepseek_ai, @MoonshotAI_Kimi, @zhipuai_cn, @minimaxai_global. Enable push notifications for each. These are the fastest English-language surfaces for all five major Chinese labs.
- GitHub Watches: Go to github.com/QwenLM/Qwen3 and github.com/deepseek-ai/DeepSeek-V3. Click Watch → Custom → Releases. GitHub emails you when a new release tag is created — this fires for point releases and major launches alike.
- HuggingFace trending: Bookmark huggingface.co/models?sort=trending. Check weekly or use a change detection tool (e.g., Visualping) set to alert on new entries from Qwen, deepseek-ai, MiniMaxAI, or THUDM.
- RadarAI weekly digest: Visit radarai.top/en/china-ai-updates each Monday for benchmark context and licensing verification on releases from the previous week.
FAQ
- How do I track new Chinese LLM releases in real time?
- Follow @qwen_lm, @deepseek_ai, @MoonshotAI_Kimi, @zhipuai_cn on X with notifications on. Watch QwenLM GitHub and deepseek-ai GitHub for release tags. Check HuggingFace trending weekly. RadarAI weekly digest provides benchmark context and licensing verification within the week.
- Where is Qwen new model release announced first?
- @qwen_lm on X (simultaneous with or minutes before GitHub), then QwenLM GitHub (technical report and model card), then qwenlm.github.io (blog with benchmark context). HuggingFace model page goes live within hours. English media follows 3–6 hours after the X announcement.
- How quickly does English media cover Chinese LLM releases?
- English media lag behind lab primary surfaces by 3–6 hours for major releases. Lab X accounts post in English within hours with key facts (model name, benchmark headline, license). English media adds market context but is not the first source for technical specifics like benchmark reproducibility or architecture details.
- Is there an RSS feed or X account for Chinese AI model releases?
- X accounts (notifications on): @qwen_lm, @deepseek_ai, @MoonshotAI_Kimi, @zhipuai_cn, @minimaxai_global. GitHub RSS: github.com/QwenLM/Qwen3/releases.atom and github.com/deepseek-ai/DeepSeek-V3/releases.atom. RadarAI weekly digest covers all major releases with context.
- What Chinese AI models were released in 2025 and 2026?
- Major releases (reverse chronological): Qwen3 series (April 2026, Apache 2.0); DeepSeek-Prover-V2 (May 2025, MIT); Kimi k1.5 (Feb 2025, proprietary); DeepSeek-R1 (Jan 2025, MIT); MiniMax-Text-01 4M context (Jan 2025, Apache 2.0); DeepSeek-V3 (Dec 2024, MIT); GLM-4-Plus (Sep 2024); Qwen2.5 series (Sep 2024, Apache 2.0); Qwen2 series (Jun 2024, Apache 2.0).
- How do I know when DeepSeek releases a new model?
- Follow @deepseek_ai on X with notifications. Watch huggingface.co/deepseek-ai for new model pages. Watch github.com/deepseek-ai for new repos. Set Google Alert for "DeepSeek" + "release". Major releases have come roughly every 3–6 months; @deepseek_ai X account is the fastest English-accessible surface.
Companion pages in this cluster
| If your question is about… | Go to | What's there |
|---|---|---|
| Full standing model watchlist with benchmarks | China AI Models List | Standing watchlist with action triggers and verification paths |
| Weekly digest of all China AI signals | China AI Updates | Weekly signal digest, curated for builders — includes all releases |
| How to access these models via API | China AI API Access Guide | Qwen, DeepSeek, Kimi API international availability and pricing |
| China AI vs US AI model benchmarks | China AI vs US AI | Dimension-by-dimension ecosystem comparison with benchmark context |
| Best English sources for China AI news | News Sources in English | Layer-by-layer routing guide for lab releases, policy, and funding |
Quotable summary: Chinese LLM release velocity in 2024–2026 has been the highest of any AI ecosystem globally: four benchmark-setting open-source releases in five months (DeepSeek-V3, DeepSeek-R1, MiniMax-Text-01, Kimi k1.5) followed by Qwen3 in April 2026. The fastest English tracking setup is X notifications from lab accounts (@qwen_lm, @deepseek_ai) plus GitHub release watches on QwenLM and deepseek-ai — these consistently beat English media by 4–6 hours on the technical details that matter for build decisions.