Changelog • LM Studio 0.3.18

Build 3

Added tool calling support for Mistral v13 tokenizer models (with proper chat templates)
Fixed the format of errors when streaming from OpenAI compatible endpoints to match the expected format

Build 2

Added support for "stream_options": { "include_usage": true } when streaming using OpenAI compatible endpoints
Upgrade @huggingface/jinja to version 0.5.0 to support more prompt templates
Fixed a bug where the app would recommend an excessively large first model for computers that can only do CPU inference
Fixed a bug where you cannot disable errored plugins
Adds Hindi and Marathi translations

Build 1

MCP: option to force restart an MCP server under the ••• menu
MCP: added a button to refresh the list of tools
Responsive UI for Mission Control on narrow windows
Show Chat screen sidebars in popovers when the app window is narrow
Added a timeout field to entries in mcp.json which can be used to specify max time for each tool call in milliseconds
Added experimental JavaScript code interpreter plugin using Deno
- Interested in plugins? Sign up for the plugins private beta here: https://forms.gle/ZPfGLMvVC6DbSRQm9
Various UI and color adjustments
Fixed a bug where chat message draft and attached files would not carry over to a newly created chat
Fixed a bug where tool use does not work with llama models (this.endToolCallStrings is not iterable)
Fixed a bug with OpenAI compatible endpoint where response_format.type did not support text
Fixed a bug where parallel tool calls that are split across multiple "chunks" were being silently dropped
Fixed a bug where SSE MCP servers would not work correctly
Fixed a bug where root level $defs were being stripped in tool definitions