Skip to content

fix: update gemini-live model to use realtime mode instead of chat#18009

Closed
Chesars wants to merge 3 commits intoBerriAI:mainfrom
Chesars:fix/remove-gemini-live-model-from-catalog
Closed

fix: update gemini-live model to use realtime mode instead of chat#18009
Chesars wants to merge 3 commits intoBerriAI:mainfrom
Chesars:fix/remove-gemini-live-model-from-catalog

Conversation

@Chesars
Copy link
Collaborator

@Chesars Chesars commented Dec 15, 2025

Title

fix: update gemini-live model to use realtime mode instead of chat

Relevant issues

N/A

Pre-Submission checklist

  • I have Added testing in the tests/litellm/ directory, Adding at least 1 test is a hard requirement - N/A (JSON config only)
  • My PR passes all unit tests on make test-unit
  • My PR's scope is as isolated as possible, it only solves 1 specific problem

Type

🐛 Bug Fix

Summary

The gemini-live-2.5-flash-preview-native-audio-09-2025 model was incorrectly configured with mode: "chat" and REST API endpoints, but this model only works with WebSockets (Realtime API).

Changes

  • Change mode from "chat" to "realtime"
  • Update supported_endpoints from ["/v1/chat/completions", "/v1/completions"] to the correct realtime endpoints:
    • gemini/ prefix: /v1/realtime
    • vertex_ai/ prefix: /vertex_ai/live

Files changed:

  • model_prices_and_context_window.json
  • litellm/model_prices_and_context_window_backup.json

@vercel
Copy link

vercel bot commented Dec 15, 2025

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Review Updated (UTC)
litellm Ready Ready Preview, Comment Dec 17, 2025 2:05am

Copy link
Member

@ishaan-jaff ishaan-jaff left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please leave it on the catalog, we support it through pass through live API requests

I suggest fixing the supported endpoints to not include /chat, /completions

The gemini-live-2.5-flash-preview-native-audio-09-2025 model only works
with WebSocket (Live API), not REST endpoints. Changed supported_endpoints
from /v1/chat/completions to /vertex_ai/live to reflect the actual
passthrough endpoint available in LiteLLM proxy.
The gemini/ prefix indicates Google AI Studio, which uses /v1/realtime
endpoint (OpenAI-compatible), not /vertex_ai/live.
@Chesars
Copy link
Collaborator Author

Chesars commented Dec 15, 2025

Please leave it on the catalog, we support it through pass through live API requests

I suggest fixing the supported endpoints to not include /chat, /completions

Thanks for the feedback! You're right - Updated the PR to fix supported_endpoints:

  • gemini-live-* (vertex_ai) → /vertex_ai/live
  • gemini/gemini-live-* (gemini) → /v1/realtime

@Chesars Chesars requested a review from ishaan-jaff December 15, 2025 23:16
@krrishdholakia
Copy link
Member

krrishdholakia commented Dec 16, 2025

shouldn't you also fix the mode of the model? as it's not a chat model. not sure what mode we have for realtime models, but i assume it's just realtime?

The mode field is used by health checks to determine the correct
check method (WebSocket for realtime vs REST for chat).
@Chesars Chesars changed the title fix: remove gemini-live model from catalog (WebSocket only) fix: update gemini-live model to use realtime mode instead of chat Jan 18, 2026
@Chesars
Copy link
Collaborator Author

Chesars commented Jan 18, 2026

shouldn't you also fix the mode of the model? as it's not a chat model. not sure what mode we have for realtime models, but i assume it's just realtime?

Exactly, updated in 763b00a

@Chesars
Copy link
Collaborator Author

Chesars commented Mar 4, 2026

Closing as superseded by #22814.

@Chesars Chesars closed this Mar 4, 2026
@Chesars Chesars deleted the fix/remove-gemini-live-model-from-catalog branch March 4, 2026 22:44
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants