Xiaohongshu — yes, the Rednote content app — just joined China’s frontier open-source club. Its dots-studio team released dots3-note preview on August 14: an open-weight 280B MoE that activates only 16B parameters per token, takes text, image, video, and audio input, handles 512K context, and ships under Apache 2.0.
The hook: a same-series checkpoint earned a certified 42/42 at IMO 2026, beating Gemini Deep Think’s 35. To be precise, that perfect score came from an internal harness, not the exact released weights — but this is still the closest anyone gets to downloading an IMO-gold model.
What you actually get
Self-reported numbers: 78.4% on SWE-bench Verified, 79.1% on MMMU Pro. The release also introduces TEMPO, an RL method for agent tasks that run tens of hours — the model checks its own progress and updates memory mid-task.
Run it yourself
Weights are on Hugging Face and ModelScope, with an FP8 variant for cheaper serving. Only 16B active means far less compute than a dense 280B. Obvious targets: self-hosted multimodal agents, long-video understanding, coding agents.
After Qwen, Kimi, and GLM, a content-community company is now shipping frontier weights. That’s the real story.
You Might Also Like
- Lm Studio Bionic Runs Claude Code Style Agents on glm 5 2 and Kimi k2 7
- Ggml Llama cpp Joins Hugging Face and Honestly it was Only a Matter of Time
- Pollen Robotics Reachy Mini a 299 Desktop Humanoid That Runs 1 7m Hugging Face Models
- Title x Square Robot Wall a With Wall oss 276m From Xiaomi 62 dof Weights on Hugging Face
- Tencent Hunyuan hy3 Preview Goes Open Source 295b moe 21b Active 256k Context
Top AI Product 
Leave a comment