Gemini 3 Flash (Vertex AI): reasoning_effort=minimal should map to thinkingLevel=minimal; current releases map to low

[joeeeeey]

Title
Gemini 3 Flash (Vertex AI): reasoning_effort="minimal" should map to thinkingLevel="minimal"; current releases map to "low"

Summary
When calling Gemini 3 Flash via Vertex AI through the LiteLLM Proxy, setting reasoning_effort="minimal" should result in thinkingConfig.thinkingLevel="minimal" (per Google’s Thinking Levels guidance for Flash). On the following LiteLLM versions, it still maps to "low":

• v1.80.0-nightly
• v1.80.10-nightly
• v1.80.10.rc.5

This prevents users from requesting the intended lowest thinking level for Gemini 3 Flash via the OpenAI-compatible chat/completions API.

Environment

• Interface: LiteLLM Proxy (OpenAI-compatible /v1/chat/completions)
• Provider: Vertex AI
• Model: vertex_ai/gemini-3-flash-preview (and gemini-3-flash)
• Configuration: reasoning_effort set in litellm_params (deployment-level), e.g.

model_list:
  - model_name: vertex_ai/gemini-3-flash-preview-fast
    litellm_params:
      model: vertex_ai/gemini-3-flash-preview
      vertex_project: <project>
      vertex_location: <location>
      vertex_credentials: <sa.json>
      reasoning_effort: minimal

Reproduction Steps

1. Start Proxy with the above deployment.
2. POST /v1/chat/completions with model="vertex_ai/gemini-3-flash-preview-fast" and a simple message payload (no need to pass reasoning_effort again in the request body).
3. Inspect the downstream request body sent to Vertex AI.

Expected

• generationConfig.thinkingConfig.thinkingLevel == "minimal" (with includeThoughts=true), for Gemini 3 Flash models (gemini-3-flash / gemini-3-flash-preview).

Actual

• On the listed versions, reasoning_effort="minimal" maps to thinkingLevel="low", and default injection also sets thinkingLevel="low" for Gemini 3 when not provided.

Why this matters

• Google documents distinct thinking levels for Gemini 3 Flash (minimal, medium, low, high). Minimal is valid and intended to minimize latency/cost for Flash.
• Mapping minimal→low prevents users from selecting the correct thinking level.

Code Pointers

• Mapping path: litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py
  • map_openai_params(...) → sets optional_params["thinkingConfig"] for Gemini 3 via _map_reasoning_effort_to_thinking_level(...)
  • _map_reasoning_effort_to_thinking_level(...) currently returns "low" for "minimal" without a Flash-specific override in the affected releases.
  • There is also a default injection that sets thinkingLevel="low" for Gemini 3 when neither thinkingLevel nor thinkingBudget are present.

Suggested Fix

1. In _map_reasoning_effort_to_thinking_level(reasoning_effort, model):

   • Detect Flash family: ("gemini-3-flash" in model.lower()) or ("gemini-3-flash-preview" in model.lower()).
   • For Flash:
     • minimal → {thinkingLevel: "minimal", includeThoughts: true}
     • medium → {thinkingLevel: "medium", includeThoughts: true}
     • disable/none → {thinkingLevel: "minimal", includeThoughts: false}
   • For Pro (e.g., gemini-3-pro-preview): keep the current behavior (minimal/medium map to low; low/high unchanged).

2. Ensure map_openai_params(...) continues to use thinkingLevel for Gemini 3 (not thinkingBudget) and does not override an explicit thinkingConfig provided by the user.

3. Consider updating docs to clarify the Flash vs Pro mapping on Gemini 3 (+ default thinkingLevel injection differences).

Request

• If this fix already exists on main, please publish a new nightly/patch release including it so Proxy users on Vertex AI can rely on reasoning_effort="minimal" → thinkingLevel="minimal" for Gemini 3 Flash.

Additional Notes

• We validated that deployment-level litellm_params.reasoning_effort is merged by the Router and reaches the Gemini handler; the discrepancy is only in the thinkingLevel mapping for Gemini 3 Flash.

[Chesars]

Hi, Thanks for reporting this! The fix is already in main. Commit 9e3f946a2e ("Add support for new flash model") added the correct mapping for reasoning_effort="minimal" → thinkingLevel="minimal" for Gemini 3 Flash. The next release should include it.

[doudoujay]

@Chesars when would we have the next release?

[Sameerlite]

@doudoujay It should be there by end of day today