The Short Answer
Frontier-grade reasoning is now downloadable. RedNote's dots studio released dots3-note-preview on August 14, 2026 under Apache 2.0 — 280B parameters, 16B active, a 512K context window, multimodal across text, image, video and audio. Weights you can download are weights you can host, and on ibl.ai you own all the code and the data, run it model-agnostic across any model, with no per-seat pricing, so you can deploy anywhere.
One clarification first, because the headline circulating is wrong in a way that matters.
dots-note-3.0 is the model that earned an officially certified perfect score of 42/42 at the 2026 International Mathematical Olympiad, which concluded on July 21, 2026 — the first AI to do so. RedNote said it would open-source that model "in the near future," without further detail. It has not shipped.
dots3-note-preview is a different model in the same series, and it is the one that shipped as open weights on August 14. Its model card makes no IMO claim at all.
The distinction is worth getting right, and the honest version is still the more interesting story.
Why does a sibling model matter more than a benchmark score?
Benchmark records move every few weeks and mostly tell you who is winning a race you are not running. Licences move rarely and tell you what you are allowed to own.
What changed on August 14 is not that a model is good at olympiad geometry. It is that a lineage demonstrably capable of that result now has a member released under Apache 2.0 — the permissive end of the licence spectrum, with no field-of-use restriction and no usage-tier gate.
A model you can download is a model you can inspect, evaluate against your own data, fine-tune, run behind your own firewall, and keep running next year at the same version. That is a different category of asset from API access, regardless of which one scores higher this month.
What are the actual specifications?
Per dots studio's own release and the published model card:
- 280B total parameters, 16B active — a Mixture-of-Experts design, so serving cost tracks the 16B active figure rather than the 280B total.
- 512K token context window.
- Multimodal input across text, image, video, and audio; text output.
- Apache License 2.0.
- TEMPO, a reinforcement-learning approach the team developed for long-horizon tasks, which they report outperforming GRPO on ARC-AGI 3.
The MoE ratio is the specification that decides whether this is deployable.
Sparse activation is what makes a 280B model serveable on infrastructure that could not host a 280B dense model, which is precisely why open-weight releases have become practical for organizations rather than only for labs.
The stated design target is long-horizon agency — tasks measured in hours, where trajectories are long and feedback is sparse. That is the failure mode most enterprise agent deployments actually hit, and it is a more useful thing to optimize for than another exam.
Does an open-weight model close the gap with frontier APIs?
For the tasks most organizations run at volume — retrieval, summarization, extraction, classification, structured output, tool calling — the gap has been narrow for a while and keeps narrowing. This release adds a long-context multimodal option to that set.
For the hardest frontier reasoning, closed models still lead, and they still get there first. That has been true through every open-weight release and it is true through this one.
The mistake is treating that as a binary. Routing every task to the most capable model available is the most common avoidable cost in enterprise AI: a frontier model doing document classification is a Bugatti in a car park.
Routing by task — open weights locally for volume and anything that cannot leave the network, a frontier API for the narrow set that genuinely needs one — is both cheaper and more robust.
We made this argument at length in Model-Agnostic AI: Why Single-Vendor Lock-In Is the Real Risk and traced the broader trend in The Open-Source Model Explosion Is Rewriting Enterprise AI Strategy.
What does "ownable frontier reasoning" actually require?
Weights alone are not a deployment. A downloadable model gets you the hardest component and none of the surrounding ones.
Running it in production requires a serving stack with continuous batching and quantization tuned to your latency target, permissions-aware retrieval so a search cannot surface a document the user could not open directly, guardrails and prompt-injection defense, evaluation harnesses built from your own traffic rather than public benchmarks, and audit logging a compliance reviewer will accept.
That is the layer ibl.ai provides, and it is the reason a licence change is genuinely actionable rather than merely interesting.
The platform is model-agnostic by construction: dots3-note-preview on your own GPUs for long-context multimodal work, a different open model for high-volume classification, a frontier API for the narrow hard cases — and the routing between them changed without touching application code.
You own all the code and the data, there is no per-seat pricing, and the same deployment runs in your cloud, on-premise, or on a fully air-gapped network.
What should an organization actually do about this release?
Three things, in order.
Evaluate it on your own data. Public benchmarks — including a perfect IMO score, which belongs to a different model in this family — predict very little about performance on your documents and your tasks. Build an evaluation set from real traffic before forming a view.
Check the licence yourself. Apache 2.0 is permissive, but "open source" is used loosely across this market and several prominent open-weight releases carry usage restrictions that matter at enterprise scale. Read the actual licence file rather than the announcement.
Decide where it runs before you decide whether to use it. Sparse activation makes a 280B model serveable, but serveable is not free. Sustained high volume favors owned GPUs; bursty or exploratory workloads favor an API. That break-even is worked through in API Inference vs Self-Hosted GPU.
The pattern to notice is not this model. It is the cadence: capability that was API-only eighteen months ago now arrives with downloadable weights within weeks of the record it set.
Any architecture that assumes a permanent frontier gap is planning against a trend that has not held. We traced the same dynamic in The Open-Weight Tipping Point.