DeepSeek-V3.2 combines reasoning with tool use

From experiment to a new release

DeepSeek announced V3.2 on December 1, 2025, as the successor to V3.2-Exp. The release described thinking integrated with tool use and introduced a separate Speciale variant for intensive reasoning evaluation.

The Speciale access route was temporary and had different tool capabilities. Historical examples using that endpoint should not be copied into a current production integration without checking the documentation.

Modes need explicit testing

A model may support tools in one mode and behave differently in another. The application should verify the complete sequence: a tool request, the tool result, and the model's next response. A successful standalone answer does not test that loop.

Repeated tool calls also need a stable way to identify work that has already completed.

Preserve the evidence trail

Log the model identifier, mode, tool name, result status, and final outcome without exposing credentials or unnecessary sensitive data. Test a failed tool and an incomplete result. Keep usage limits around the full task, not only each individual request. A reasoning-and-tools integration becomes useful when it can recover predictably and show an operator why a task succeeded or stopped.