I’m looking at running a DeepSeek model locally and would like it to call tools such as a search function or a small internal API. Some serving stacks expose an OpenAI-compatible endpoint, but I’m not sure whether that means the model’s tool calls are genuinely supported or just formatted through prompting.

For people who have tested native tool calling with DeepSeek, which model and inference setup did you use? I’m interested in how reliably it chooses a tool and returns arguments that the host application can parse.