Changelog 👾

Build 2

• Fixed a bug where the parameters reported back in /v1/responses API are sometimes incorrect
• Fixed a bug where input_tokens and cached_tokens were sometimes reported incorrectly in /v1/responses API
• /v1/responses API now return better formatted errors

Build 1

• Support for image_url input in OpenAI-compatible /v1/chat/completions REST endpoint
• Support for top_logprobs in OpenAI-compatible /v1/responses REST endpoint
• Output cached_tokens statistics in OpenAI-compatible /v1/responses REST endpoint

Build 1

• [Mac][M5] Enable auto-upgrade to MLX NAX engine to fix MLX model crashes and improve performance

Build 1

• Support for the LFM2 tool call format
• Fix "Cannot read properties of null (reading 'architecture')" when using a generator

Build 1

• FunctionGemma Support

Build 1

• [MLX] Support for Devstral-2 and GLM-4.6V
• Fixed a bug where the default system prompt was still sent to the model even after the system prompt field was cleared.
• Fixed a bug where exported chats did not include the correct system prompt.
• Fixed a bug where the token count was incorrect when a default system prompt existed but the system prompt field was cleared.
• Fixed a bug where sometimes the tool call results were not being added to the context correctly

Build 1

• Support for EssentialAI's rnj-1 model
• Fix jinja prompt formatting bug for some models where EOS tokens were not being included properly

Build 1

• Support for MistralAI's Ministral models (3B, 8B, 13B)
• Support for Olmo-3 tool calling

Build 2

• Support for GLM 4.5 tool calling
• [MLX] Fixed prompt template bug that caused GLM-4.1V to not recognize images
• Support for olmOCR-2
• Fix a bug where sometimes Download button would continue showing for an already downloaded model

Build 1

• Support for passing base64 images into OpenAI-compatible /v1/responses endpoint
  • See https://platform.openai.com/docs/guides/images-vision?api-mode=responses&format=base64-encoded#giving-a-model-images-as-input for details
• Flash Attention is now enabled by default for Vulkan and Metal llama.cpp engines
• Fix OpenAI-compatible /v1/responses endpoint "previous_response_not_found" bug due to internal file read error
• Fix a bug where update toast would sometimes retrigger and close
• Fix the "No model selected for this chat and no lastUsedModel recorded. Please select a model" error
• Fixed cases where downloading additional variants of the same model sometime wouldn't get nested correctly

New in LM Studio 0.3.31:

• Improve image understanding output quality, especially for OCR tasks
• Default image attachment size is now 2048px on the longest side for better quality with vision models
  • Can be changed in Settings > Chat > Image Inputs
• Support for MiniMax M2 tool calling
• Flash Attention is now enabled by default for CUDA engines
• New CLI commands to manage engines:
  • lms runtime get
  • lms runtime update
• Improved macOS 26 compatibility and support

Build 7

• Fixed a bug where sometimes the Load Model button was greyed out for large MLX models
• Set Flash Attention on by default for CUDA engines

Build 6

• Added tool use support for MiniMax M2
• Added better controls for image input size for vision models

Build 5

• Update dependencies to support macOS 26
• Server UX improvements
  • Log v1/chat/completions prompt processing progress to Developer Logs
  • Model-specific v1/chat/completions Developer Logs
  • Log error when JIT model load failed due to guardrails settings

Build 4

• Fixed "vision.imageResizeSettings key does not exist" error
• Added settings for configuring maximum image input dimensions for vision models.
  • Accessible via Settings > Chat > Image Inputs > "Image resize bounds"
  • These settings control the maximum dimensions to which input images are resized before being sent to vision-capable models.

Build 3

• Added lms runtime get and lms runtime update CLI commands to manage runtime extensions from the terminal.
  • Run lms runtime -h for more info.
• Fix issue where sometimes reasoning models' output would start with a plaintext <think> token

Build 2

• Increase image resize limits to 1024x1024 for improved image processing performance (superseded: now 2048px)
  • In a future update this will be user configurable (fixed in build 4)

Build 1

• [MLX] Fix ValueError: Image features and image tokens do not match for Qwen3 VL Thinking models
• Fix occasional UI crash when searching models

Build 2

• [MLX] Fix ValueError: Image features and image tokens do not match for Qwen3 VL Thinking models
• Fix occasional UI crash when searching models

Build 1

• Fixed streaming mode bug that was impacting tool calling functionality for Qwen 3 Coder in /v1/chat/completions API.
• [Vulkan] Fix bug where models were not being loaded onto iGPUs (requires runtime update)
• Compatibility support for the developer role in /v1/responses API endpoint. For now, developer messages will be processed as system messages internally.