DeepSeek V3.2 Brings Olympiad-Level Reasoning to Open Weights Under MIT Licence

On 1 December 2025 the Chinese lab DeepSeek released DeepSeek-V3.2 and a research variant, DeepSeek-V3.2-Speciale, both published as open-weight models under the permissive MIT licence. At roughly 685 billion parameters and about 690 GB on disk, they are the same scale as the V3 family that shook markets in January 2025, but with a training recipe focused on reasoning and agentic behaviour. V3.2 became the default model on DeepSeek's chat service and API on the day of release.

According to DeepSeek's model documentation, V3.2 blends reasoning, agent and human-alignment data distilled from specialised models and then undergoes thousands of steps of continued reinforcement learning. The Speciale variant was trained only on reasoning data with a reduced length penalty, allowing it to think longer, and was augmented with mathematical datasets aimed at formal proof. The company reports that both models reached gold-medal level on the 2025 International Mathematical Olympiad, a bar that until mid-2025 was achieved only by unreleased systems from Google DeepMind and OpenAI. Independent evaluators have broadly confirmed strong mathematics and coding results while noting that the models trail Gemini 3 Pro and GPT-5-class systems on some agentic and multimodal tasks.

The release builds on the experimental V3.2-Exp published in September 2025, which introduced DeepSeek Sparse Attention to cut the cost of long-context inference and was accompanied by a steep API price cut. The December models make that architecture the production default, keeping DeepSeek's position as the lowest-cost provider of frontier-adjacent capability.

The strategic significance lies less in any single benchmark than in the trajectory. In eleven months the open-weight frontier has moved from "surprisingly good" to "olympiad gold", and the licence terms have become more permissive rather than less. For enterprises this widens the realistic option set: self-hosted or sovereign-cloud deployments of a model at this level are now feasible, at the cost of taking on evaluation, safety tuning and infrastructure responsibilities that closed-model vendors otherwise absorb. It also complicates policy. US export controls on advanced chips were meant to slow exactly this kind of progress; DeepSeek's continued cadence suggests algorithmic efficiency is partly offsetting hardware constraints.

There are governance caveats. DeepSeek publishes limited information about training data, safety evaluations or red-teaming, and its hosted service is subject to Chinese data and content law, which several Western regulators have already flagged. Open weights allow independent scrutiny, but they also mean any safety behaviour can be removed by fine-tuning. Under the EU AI Act, a model of this scale falls within general-purpose AI obligations, and downstream deployers inherit documentation and transparency duties whether or not the upstream provider engages with the AI Office.

What it means for leaders

• Distinguish the weights from the service. Self-hosting V3.2 on your own infrastructure carries a very different data-sovereignty profile from calling DeepSeek's API; policy should address the two separately.
• Budget for evaluation you would otherwise outsource. With open weights, your organisation owns bias, safety and robustness testing. Use NIST AI RMF's Measure function as the checklist.
• Verify licence compatibility and provenance. MIT terms are permissive, but confirm the model card, checksums and any usage restrictions before the weights enter a regulated product.
• Plan for GPAI obligations downstream. If you build on an open model whose provider offers little documentation, you will need to compensate in your own technical file and transparency notices.
• Use the cost curve deliberately. Open olympiad-level reasoning at commodity prices is a negotiating lever with closed-model vendors and a reason to revisit which workloads truly need a frontier API.