Dots3-Note Preview: RedNote's 280B Apache-Licensed Omni Model

Most open-weight models that accept images, audio and video are either small enough to be clearly second-tier or large enough that only a cluster can host them. Dots3-Note Preview is an attempt at the middle: 280B total parameters, 16B active, a 512K-token window, and an FP8 checkpoint that its makers say fits on a single eight-GPU node.

It comes from dots studio, the AI lab of Xiaohongshu, the Chinese social platform known in English as RedNote. It is less documented than the big-lab releases, it is labelled a preview, and its full technical report has not yet been published. All of that shapes how you should approach it.

Who makes it, and what "preview" means here

The lab's earlier open release, dots.llm1, was a text-only mixture-of-experts model. Dots3-Note Preview is described as the lightest model in a planned Dots3 family, which implies larger siblings to come. OpenRouter lists its release as 14 August 2026.

The preview label is concrete rather than decorative. The model card says the full report is "coming soon". At the time it was written, Transformers support depended on an unmerged pull request, SGLang support came through a dedicated development image, and prefill CUDA graphs were not yet supported. Expect the serving story to change, and expect a non-preview release to behave differently from this one.

What is inside the 280B

The model card is more specific about architecture than many larger releases:

• Language model: a mixture-of-experts with 256 routed experts plus one shared, eight active per token, about 16B activated out of 280B total.
• Attention: a hybrid of 13 DeepSeek Sparse Attention layers and 33 sliding-window attention layers.
• Vision encoder: itself a mixture-of-experts ViT, 7B total with 1.2B activated.
• Audio encoder: a dense 800M model.
• Vocabulary: about 152K tokens.

The attention mix is the interesting design choice. Sliding-window layers only look at recent tokens and are cheap; sparse attention layers select relevant tokens across the whole context. Mostly-local attention with a minority of global layers is a common way to make long windows affordable, and it is how a 16B-active model can offer 512K tokens. It is also exactly the design where quality deep inside a long context needs checking rather than assuming. Long context degradation covers what to look for.

One small inconsistency: the card text says 280B, while the checkpoint metadata reports about 288B. The difference is probably the encoders, but it is a reminder to check which figure a comparison is using.

The benchmark claims, and the asterisks

The model card lists these scores:

• SWE-bench Verified: 78.4
• SWE-bench Pro: 61
• SWE-bench Multilingual: 75.7
• Terminal-Bench 2.1: 75.1
• Toolathlon (Verified): 55.6
• MMMU-Pro: 79.1

The headline charts are images, and the details sit in the appendix footnotes, which are worth reading before quoting any row. Terminal-Bench 2.1 was run on the Terminus-2 harness at temperature 0.7 with a 256K window; the SWE-bench rows used live-swe-agent with a tailored instruction prompt; APEX-Agent (30.8) was scored in a multimodal setting with Gemini 3.5 Flash as the judge; and the HLE figure (52.6) is with tools, not closed-book. Asterisks in the tables mark scores the lab measured itself rather than took from official leaderboards, mostly for competitor models. Each of those choices can move a score by several points, so treat the figures as indicative until the full report is out.

If the numbers hold, they are strong for 16B active. A SWE-bench Pro score around 61 would sit close to GLM-5.2, which reports 62.1 with roughly 40B active, and that is precisely why it is worth verifying on your own tasks rather than accepting. Vendor-claimed vs independent benchmarks covers how far to discount a table like this.

Modalities: check what your endpoint accepts

The model card lists text, image, video and audio input, with text output; video inputs include their audio track when one is present. Hosted endpoints may expose less than that. OpenRouter's catalogue lists text and image input for its free endpoint, for example, so do not design an audio pipeline around a hosted preview until you have sent it audio and had a sensible answer back.

That free endpoint is useful for evaluation and little else. Free variants are rate-limited, their terms and limits change, and they can be withdrawn. Note too that catalogues list a maximum output close to the whole context window; ignore that figure and size max_tokens to what your task needs. Free tier strategies covers how to get real answers from free access without building on it.

Apache 2.0 and self-hosting

The weights and modelling code are licensed under Apache 2.0. That is permissive in the way that matters commercially: you can fine-tune, redistribute and serve it in a paid product without a revenue threshold or a separate agreement. Apache 2.0 also includes an explicit patent grant, which MIT does not. Open-weight model licences compared covers the practical differences.

For serving, the card recommends the FP8 checkpoint on one eight-GPU node, with BF16 needing more memory. vLLM support landed on its main branch, and multi-token-prediction speculative decoding is available, which the card says can cut time per output token by more than half. Community GGUF quantisations exist for llama.cpp-based runtimes, but quantising a mixture-of-experts model with an attached vision and audio stack is where quality loss hides, so test them separately. The recommended sampling is temperature 1.0 and top_p 0.95, with enable_thinking toggling reasoning.

How to evaluate it in an afternoon

1. Pick fifteen tasks from your own work: five coding, five agentic with tools, five involving images or documents.
2. Run them on the free endpoint with thinking on and off, and on one model you already trust.
3. Include at least three prompts above 200K tokens with the key fact placed in the middle, to test the sliding-window design.
4. Score completion and correctness blind, then decide whether it is worth a self-hosted trial.

A preview from a smaller lab can be a real find or a benchmark artefact, and an afternoon on your own tasks will tell you which faster than any leaderboard. How to benchmark LLMs on your own work has the fuller method.