Reasoning

DeepSeek’s official API turns on chain-of-thought with reasoning_effort or thinking: { "type": "enabled" } and returns a separate reasoning_content field.

Gonka DeepSeek hosts use an enable_thinking switch instead. Without translation, reasoning tokens often land in content, which breaks tool-call parsers.

The setting

In Settings enable Translate reasoning_effort / thinking for DeepSeek.

• Off by default
• Only applies to DeepSeek model ids
• Does not rewrite requests that already set enable_thinking or chat_template_kwargs.thinking
• Cannot reproduce low / medium / high / max budgets — vLLM is on/off, not a token budget

What clients should send

Keep sending the official fields. With the toggle on, the proxy maps them for the host and splits reasoning_content back out.

Related

• Tool calling
• Settings
• Available models