RedNote Scored a Perfect 42/42 at the IMO. You Get the Sibling.

TL;DR

On August 14, dots studio, the model lab inside Chinese social platform Xiaohongshu (RedNote), open-sourced dots3-note preview under Apache 2.0. It is a 280B-parameter mixture-of-experts model with 16B active, a 512K context window, and native text, image, video and audio input. It is the first open release from the dots3 family. It is also not the family member that scored a perfect 42 out of 42 at this year's International Mathematical Olympiad. That one is called dots-note-3.0, it is still closed, and yes, the two names differ by the position of a single character.

What actually shipped

The weights are on Hugging Face and ModelScope, Apache 2.0, no revenue-share rider, no acceptable-use annex. The model card is blunt about where this sits in the lineup: it is the first open-weight model in the dots3 family and the most lightweight member of that family, in the sense that a 280-billion-parameter checkpoint can be described as lightweight.

The architecture is one dense layer followed by 45 MoE layers, 256 routed experts plus one shared expert, top-8 routing, hidden size 5120, and a 152K vocabulary. Multimodality is bolted on with real encoders rather than a caption pipeline: a MoE vision transformer at 7B total and 1.2B active, plus a dense 800M audio encoder. Input takes text, image, video and audio. Output is text only.

That total-versus-active split is the whole economic story of an MoE, and it is the part people consistently get wrong when they budget hardware. Think of a hospital that keeps 256 specialists on payroll: you pay rent and salary for all of them every single day, but any one patient gets routed to eight. Your VRAM bill is the payroll. Your tokens per second is the appointment.

The IMO trophy belongs to a different model

Last month the lab took the first perfect score in IMO history, 42 out of 42, on officially graded contest problems. SCMP reported it at the time, alongside the useful baseline that at the previous year's IMO both Google DeepMind and OpenAI landed 35 out of 42, which was already gold-medal territory.

That run was credited to dots-note-3.0. The thing you can download is dots3-note preview, which 36Kr describes as a counterpart in the same series tuned for long-horizon tasks, explicitly not the IMO model. So the headline number and the downloadable artifact are related the way a sibling is related, not the way a copy is.

The numbers, and where they stop flattering

Every figure below is vendor-reported from the model card, which means it has had exactly as much independent scrutiny as any other launch-day table: none. Read it as a claim, not a result.

The coding lane is where it looks strongest. SWE-bench Verified comes in at 78.4, SWE-bench Multilingual at 75.7, and the harder SWE-bench Pro at 61. On the multimodal side, MMMU Pro reports 79.1 on the ten-option standard split. Humanity's Last Exam sits at 52.6.

Then there is Apex Agents at 30.8, which is the number the marketing does not lead with, and the one you should read first if you plan to point this at a long-horizon agent loop.

There is one third-party-ish datapoint worth noting. On ARC-AGI-3, 36Kr reports the model around 0.35 at a test-time cost under $500, below Claude Opus 4.8 on high effort in absolute score but far ahead of it on cost per point. That is the shape of most strong open-weight releases right now: not the top of the chart, considerably better value on the way up it.

What it takes to run

The recommended deployment is a single node of eight H100s with FP8 weights. BF16 is also supported, which roughly doubles your problem. At FP8 the weights alone are on the order of 280 GB before you have allocated a single token of KV cache, and a 512K context window is not shy about KV cache.

Serving is vLLM with native support, plus an SGLang path via a dedicated Docker image, with tensor parallel and expert parallel options exposed. Transformers works for local poking. Note that the model card flags some framework support as still pending upstream PR merges, so pin your versions and expect the first week to be the first week.

If you want to try before provisioning anything, the preview is listed free on OpenRouter.

Why the license is the actual news

This has been a loud fortnight for open weights, and a mixed one. The current wave has produced plenty of releases where the weights are public and the license is a negotiation, including flagship checkpoints landing under bespoke revenue-share terms rather than a recognized open license. Against that backdrop, a 280B omni-modal model arriving under plain Apache 2.0, no strings, is the part a builder should actually care about.

It also arrives two days after Qwen3.8-27B, into a week where the open-weight release calendar has been busier than the closed one. Not a bad problem to have.

Caveats before you rebuild your stack around it

• Preview means preview. The name is not decoration, and the framework support is explicitly still landing.
• Every score here is self-reported. No neutral lab has rerun the table yet. Treat the coding numbers as a hypothesis you test on your own repo.
• The IMO result is not this model's. It is real and it was officially graded, but it belongs to the closed sibling. Do not let it set your expectations for the download.
• Agent performance is the soft spot. An Apex Agents score of 30.8 next to a SWE-bench Verified score of 78.4 is a model that fixes issues well and plans long chains less well.
• This is not a homelab model. Eight H100s is the recommendation, not the ceiling.

Key Takeaways

• dots studio open-sourced dots3-note preview on August 14 under Apache 2.0: 280B total, 16B active, 512K context, text plus image plus video plus audio in, text out.
• The 42 out of 42 IMO score belongs to dots-note-3.0, a different and still-closed model in the same family.
• Vendor-reported highlights: SWE-bench Verified 78.4, MMMU Pro 79.1, Humanity's Last Exam 52.6, and Apex Agents 30.8.
• Recommended serving is eight H100s at FP8 via vLLM or SGLang, with a free preview on OpenRouter if you just want to kick the tires.
• The genuinely notable part is the license: an unencumbered Apache 2.0 release at this scale, in a season of bespoke open-ish terms.

Sources: dots-studio/dots3-note-prev model card (Hugging Face), dots3-note-prev repository (GitHub), 36Kr, South China Morning Post, OpenRouter.