dots3-note preview
RedNote's first open dots3: 280B omni-modal in, 512K context — with an Apache LICENSE exactly where it belongs
What it is
The first open-weights release in Xiaohongshu's dots3 family — the sibling line of the lab's IMO perfect-score model. A 280B-total / 16B-active MoE that takes text, images, video, and audio in and writes text out, through a 7B MoE ViT vision encoder (1.2B active) and an 800M audio encoder, with 512K tokens of context. Built, per dots studio, for 'long-term daily tasks' — the days-long agent runs its self-evaluation loop targets — and the agentic numbers land the claim: 78.4 SWE-bench Verified and 75.1 Terminal-Bench 2.1, self-reported.
Why it's interesting
This sat on our watch list as 'Apache-2.0 reported, verify on the official org' — and it verifies cleanly twice over. The official GitHub repo's LICENSE is the full standard Apache-2.0 text (11,357 bytes, read from a fresh clone on 2026-08-31), the README states 'the dots3-note preview model weights are released under the Apache License 2.0,' and the HF weights repo carries its own LICENSE file with the apache-2.0 tag. Copyright Xiaohongshu, released by dots studio, no rider anywhere we could find. A 280B omni-modal MoE under clean Apache is the kind of release the fine-print shelf makes you stop taking for granted.
Use cases
- Long-horizon agents that watch, listen, and read — omni-modal input on one self-hosted stack
- 512K-context document, video, and audio pipelines without per-token API anxiety
- Fine-tuning an Apache-licensed current-gen omni-modal base — no negotiation, no thresholds
Who it's for
Agent-infra teams with a real GPU node; researchers who wanted MOSS-class omni input with a clean license
Setup
Advanced. Single 8-GPU node recommended for FP8 serving (BF16 wants more); vLLM support on main, Transformers and SGLang PRs in review; free hosted access via OpenRouter to try before you download
Limitations & cautions
It's a preview checkpoint the lab says will evolve, benchmarks are self-reported with independent evals scarce, framework support is still landing (two of three integrations were open PRs at check), and the 8-GPU FP8 floor keeps it out of hobbyist reach regardless of the license's generosity.
Editorial takeaway
The social-media company shipped what the model labs keep promising: frontier-adjacent omni-modal weights, a real LICENSE file, and no fine print. Watch list, resolved — the way we wish they all resolved.