deepseek-v3.2

Overview. DeepSeek-V3.2 is a text model from DeepSeek, released in December 2025 under the MIT licence. It is a sparse Mixture-of-Experts model with 685 billion parameters in total and about 37 billion active per token. It accepts up to 163,840 tokens of context and can return up to 65,536 tokens. One model covers both modes: it answers directly, or it reasons first when you set a reasoning effort.

Key strengths. - Strong on agentic and coding benchmarks for an open-weight model - Reasoning on demand, with adjustable effort - Tool calling and structured output - Open weights under MIT, so the same model can be served by several providers

When to use it. Use it for agentic work and coding where the model has to call tools, return structured output or work through a task in several steps. It also suits long-document analysis. If a task does not need deliberate reasoning, set the reasoning effort to none to cut latency and output cost.

Limitations. The model accepts text only, with no image or audio input. Reasoning adds output tokens and latency, so higher effort costs more per request. The reasoning levels each endpoint accepts can differ.

Sovereignty note. The model is developed by DeepSeek, based in China, and published as open weights. The weights can be run anywhere, so the creator never sees your prompts. Where your data is processed, and under which retention and training terms, depends on the endpoint that serves the request, not on the model.