dots3-note preview: open multimodal weights, 78.4 on SWE-bench
Xiaohongshu's Dots Lab ships a 280B/16B-active MoE under Apache 2.0: text, image, video and audio in, 512K context, free on OpenRouter now.
Copy markdown
An Apache-2.0 MoE you can actually self-host
280B total but only 16B active, so it serves like a mid-size model rather than a 280B monster. Weights are on Hugging Face and ModelScope with an FP8 checkpoint that fits one 8-GPU node under vLLM or SGLang.
One model for text, image, video, and audio
A MoE vision encoder and an 800M audio encoder let it take images, video, and audio alongside text: one endpoint that can read a screenshot, watch a clip, and transcribe speech, then reason over all of it. Output is text.
78.4 on SWE-bench Verified, frontier territory for open weights
It posts 78.4 on SWE-bench Verified, 75.7 on SWE-bench Multilingual, and 79.1 on MMMU-Pro for vision. Those are real coding-agent numbers from a model you can download and run yourself.
512K context, built for long-horizon agents
This preview is the series' agent-focused build, trained with a new RL method the lab calls TEMPO for long-horizon tasks, and it carries a 512K-token window: enough for a whole repo plus a long run of tool calls.
Free to wire in on OpenRouter tonight
Before you pull hundreds of gigs of weights, the full model is live for free on OpenRouter at the same 512K context. Point your agent at it and see if it holds up on your own tasks.
Elsewhere: the IMO-perfect sibling stays closed
This preview is the open cousin of dots-note-3.0, the flagship the lab says reportedly logged a verified 42/42 at July's IMO, a first for AI. That top model isn't open; this agent build is what you get to keep.