v1.82.2-nightly
📦 litellmView on GitHub →
✨ 19 features🐛 27 fixes🔧 20 symbols
Summary
This release focuses heavily on expanding support for various models (Gemini, Mistral Voxtral, Qwen3.5) and enhancing the Multi-Cloud Proxy (MCP) features, including new authentication and UI workflows. Numerous bug fixes address streaming issues, provider specific mappings, and security concerns.
Migration Steps
- Migrate user management buttons from Tremor to Ant Design (UI change).
- Revert 'policy builder' implementation.
- Revert 'feat: add model_cost aliases expansion support' (Note: it was re-added later, this step addresses the temporary revert).
✨ New Features
- Auto-inject prompt caching support for Gemini models.
- Support detail parameter for image resolution on Gemini 2.x models.
- Add Voxtral audio transcription support for Mistral.
- Add litellm.acount_tokens() public API + OpenAI token counting support.
- Add image generation and image edit support for black_forest_labs.
- Add token authentication support for MCP servers.
- Add PANW Prisma AIRS guardrail with apply_guardrail support.
- Introduce policy builder functionality (later reverted).
- Add BYOM — non-admin MCP server submission + admin review workflow.
- Implement OpenAPI MCP server support with popular API quick-picker.
- Add missing supports_web_search to OpenAI models.
- Support input_fidelity parameter for image edit API.
- Add model_cost aliases expansion support.
- Add MCP OAuth2 value prop and server auth badges to chat-ui.
- Show MCP tools per server in chat panel in UI.
- Add Responses API support to JSON providers (openai_like).
- Add Qwen3.5 series support for OpenRouter.
- Add support for Gemini embedding 2 preview.
- Implement chat-ui feature for responses API + MCP tool execution in /chat.
🐛 Bug Fixes
- Avoid migration hook serviceaccount dependency cycle.
- Set budget_reset_at when creating customer with budget_duration.
- Map reasoning to reasoning_content in Delta for gpt-oss providers during streaming.
- Fix model management page UI issues.
- Fix user specified async client being ignored with Gemini streaming+async.
- Add missing indexes for top CPU-consuming queries.
- Fix Claude code req traces on Langfuse.
- Fix Presidio edge case with Anthropic handle on PII token leak.
- Map unknown finish_reason values to finish_reason_unspecified in stream_chunk_builder to prevent ValidationError.
- Strip secret_fields from guardrail logging response for security.
- Fix Anthropic prompt caching returning a string instead of an int.
- Fix tag usage cost panel showing zero.
- Don't auto-detect M2M OAuth from field presence for MCP.
- Add AWS SigV4 auth for Bedrock AgentCore MCP servers.
- Transform tool_choice string to object format for Snowflake.
- Fix OpenAPI OAuth flow for MCP (transport mapping, error messages, discovery bypass).
- Stop health checks triggering on server deletion for MCP Servers UI.
- Fix MCP Key Scope Type.
- Deduplicate tool_result messages by tool_call_id for Anthropic.
- Enforce type:"object" on tool input schemas in Anthropic's _map_tool_helper.
- Fix streaming empty chunk bug for providers using BaseLLMHTTPHandler (xai).
- Fix proxy startup failing if prisma migrate fails.
- Preserve thinking block order with multiple web searches.
- Unify finish_reason mapping to OpenAI-compatible values for completion.
- Fix OpenAPI OAuth flow for MCP.
- Fix MCP Key Scope Type.
- Fix proxy preserving multipart/form-data boundary in passthrough endpoints.