RedNote's IMO-perfect lab open-sources dots3-note under Apache 2.0
A 280B/16B multimodal MoE — text, image, video, audio — with 512K context and strong coding scores, free on OpenRouter and Apache-2.0 for shipping.
Copy markdown
The first open weights from dots studio
RedNote's dots studio released dots3-note-prev, a 280B-parameter mixture-of-experts with just 16B active — its first open model, under Apache 2.0. The weights are on Hugging Face in BF16 and FP8, so you can self-host, fine-tune, and ship it in a commercial product with no license strings.
The lab that just aced the Math Olympiad
dots3-note shares a series with dots-note-3.0, which RedNote says earned an officially verified perfect 42/42 at the 2026 International Mathematical Olympiad — the first AI to do so, per the South China Morning Post. The takeaway: Xiaohongshu, China's lifestyle-social app, is now a serious reasoning-model lab.
Frontier-class coding, self-hostable
The preview posts SWE-bench Verified 78.4 and MMMU-Pro 79.1 — coding and multimodal scores in the closed frontier's range, but on weights you own. It takes text, image, video and audio in at a 512K context, so one model can read a whole repo plus screenshots in a single pass.
Free to wire in this weekend
A free tier is live on OpenRouter (dots-studio/dots-3-note-preview:free) at the full 512K window, with a hosted API at dots.ai if you'd rather not run GPUs. Weekend build: a long-context agent that ingests PDFs, screenshots and video frames together and reasons over all of it.
Elsewhere: Microsoft ships its first reasoning model
Microsoft launched MAI-Thinking-1 — its first in-house reasoning model, a ~1T MoE with 35B active and a 256K context that scores AIME 2025 97% and matches Claude Opus 4.6 on SWE-Bench Pro — in public preview on Microsoft Foundry. Sibling MAI-Code-1.1-Flash is now built into GitHub Copilot and VS Code, so paid Copilot users can pick a Microsoft coder today.