Summary

Add a custom Xiaomi provider profile class that emits extra_body.thinking to control MiMo thinking mode. When the user configures reasoning_effort: none or otherwise disables reasoning, sends thinking: {type: "disabled"} to prevent unnecessary token consumption.

Closes #27325

Problem

Xiaomi MiMo provider used a bare ProviderProfile with no overrides — the thinking parameter was never passed in API requests. MiMo defaults to thinking=enabled, consuming tokens on every request regardless of user config. Settings like reasoning_effort: none had no effect on the actual API call.

Solution

Replace the bare ProviderProfile with a XiaomiProfile subclass that overrides build_api_kwargs_extras, following the same pattern used by KimiProfile:

• When reasoning_config.enabled is False → extra_body.thinking = {type: "disabled"}
• Otherwise → extra_body.thinking = {type: "enabled"}

Only 1 file changed (+45/-2).

Files Changed

Test Results

94 passed (related provider profile + Xiaomi tests)
0 regressions

Design Decisions

• Follows Kimi pattern exactly — Kimi already uses the same extra_body.thinking approach via KimiProfile.build_api_kwargs_extras. Xiaomi uses the same OpenAI-compatible API format.
• Minimal diff — no changes to transport, no new params, no legacy path changes. Only the provider profile needed updating.