DeepSeek V3.2: Open Weights, Agent Tools & Gold Claims

One weight release, two different jobs

DeepSeek shipped two models on December 1, 2025: DeepSeek-V3.2, available through its app, web chat and API, and DeepSeek-V3.2-Speciale, an API-only variant. Both were released as open weights under the MIT license, so the weights can be reused, modified and redistributed with no royalty and minimal restriction. [1][2]

DeepSeek describes V3.2 as its first model to integrate reasoning directly into tool-use, supporting tool calls in both a thinking and a non-thinking mode, and built for agent workloads. Speciale drops tool-calling entirely and is aimed only at deep reasoning tasks, the split the event's title and summary describe as agent work versus pure reasoning. [1][2]

Gold-level claims, stated by the vendor

DeepSeek's own release notes state that Speciale reached gold-level results at the 2025 International Mathematical Olympiad, the China Mathematical Olympiad, the ICPC World Finals and the International Olympiad in Informatics, and that its reasoning rivals Gemini-3.0-Pro. The company also describes V3.2 itself as reaching GPT-5-level performance. [1]

Atlas interpretation: These are self-reported scores from the model's developer, not an independent competition result verified by the olympiad organizers or a third-party evaluator. Gold-medal language describes where a model's output would have ranked against past human competitors, not a certified placement in the actual event. [1]

Sparse attention, not a new parameter count

The Hugging Face model card lists 685 billion total parameters and DeepSeek Sparse Attention, abbreviated DSA, as the release's central mechanism. DSA computes attention over a selected subset of tokens rather than the full sequence, which DeepSeek's earlier V3.2-Exp documentation credits with cutting the cost of long-context training and inference. [2]

Atlas interpretation: V3.2 and Speciale continue the sparse-attention architecture DeepSeek introduced with the experimental V3.2-Exp release months earlier rather than debuting it. The December release is a production version of that mechanism paired with new agent training data, not a change to the underlying model family that began with V3 in December 2024. [2]

One release in a fast cadence

Atlas interpretation: DeepSeek followed this release with V4 in April 2026, continuing a pattern of shipping open-weight checkpoints on a roughly quarterly cadence rather than treating any single release as a settled endpoint. V3.2's MIT license kept that pattern free of the licensing friction that limits reuse of some competing open-weight models. [1][2]