RedNote open-sources Dots3-Note: 280B MoE, Apache 2.0, free today
From the series that became the first AI to score a verified 42/42 at the IMO — now open on Hugging Face, multimodal, and cheaper than Opus 4.8 on ARC-AGI.
Copy markdown
The first AI to ace the IMO — now you can run it
dots3-note is the open preview of the dots-note-3.0 series, which became the first model ever to earn an officially verified full 42/42 at the International Mathematical Olympiad. On reasoning and agent tasks it matches or beats models several times its size.
280B total, 16B active — one 8-GPU node
A Mixture-of-Experts with 256 routed experts (top-8) and a 512K context, but only 16B params fire per token. Serve the FP8 checkpoint on a single 8-GPU node via vLLM or SGLang — frontier-class output at mid-size inference cost.
Apache 2.0 — you can actually ship it
Weights are live on Hugging Face and GitHub under a clean Apache 2.0 license: commercial use, fine-tunes and redistribution all allowed. Rare freedom for a model at this capability tier.
Free on OpenRouter today
The preview is a $0 endpoint (dots-3-note-preview:free) with the full 512K window. Point an agent at it and test frontier math and reasoning without spending a token of budget.
Multimodal in, and cheap on ARC-AGI
It takes text, images, video and audio (text out) through a 7B MoE vision encoder and an 800M audio encoder. It hits ~0.35 on ARC-AGI-3 for under $500 — more cost-efficient than Claude Opus 4.8.
Elsewhere: Zed opens Delta for coding agents
Zed's new Delta puts humans and AI agents in one shared thread where review comments stay anchored to code as it changes, and it syncs with Claude Code and other harnesses. Private beta now, signups rolling out.