← Back to Blog

LM Studio 0.3.23

By
LM Studio Team

•

2025-08-12

📣 ICYMI: LM Studio is now free for work! Read more https://lmstudio.ai/blog/free-for-work

LM Studio 0.3.23 is now available as a stable release. This build focuses on reliability improvements, performance for underpowered devices, and a handful of bug fixes.

Improve openai/gpt-oss in-chat tool calling reliability

Tool names are now consistently formatted before being sent to the model. Previously, tools with spaces in their names would confuse gpt-oss and lead to tool call failures. Tool names are now converted to snake_case.

Additionally, we squashed a few parsing bugs that could have previously led to parsing errors in the chat. You might notice significant improvements in tool calling reliability.

Reasoning Content from gpt-oss using the Chat Completions endpoint

This is a change in behavior compared with version 0.3.22.

  • message.content will no longer include reasoning content or <think> tags.
  • Reasoning is now in choices.message.reasoning (non-streaming) and choices.delta.reasoning (streaming).
  • This matches the behavior of o3-mini.

Force MoE expert weights onto CPU or GPU

In this version, we added an advanced model load setting to place all MoE expert weights onto the CPU or GPU (default).

undefined

Turn on to force MoE expert weights on CPU. Try this for low VRAM machines

This is beneficial if you don't have enough VRAM to offload the entire model to GPU dedicated memory. If this is the case, try to turn on the "Force Model Expert Weights onto CPU" option in the advanced load settings.

If you can offload the entire model to GPU memory, you're better off sticking with placing expert weights onto the GPU as well (this is the default option).

This is utilizing the same underlying technology as llama.cpp's --n-cpu-moe.

Recall that you can set persistent per-model settings. See the docs for more info.

LM Studio 0.3.23 - Full Release Notes

Upgrade via in-app update, or from https://lmstudio.ai/download.

Build 3

  • [llama.cpp][MoE] Add ability to offload expert weights to CPU/GPU RAM via "Force Model Expert Weights onto CPU" in advanced load settings
  • Tool names are normalized before being provided to the model (replace whitespaces, special chars)

Build 2

  • Fix "Complete Download" button sometimes not working when downloading a staff-picked model
  • Fix "Fix" button not working for extension packs (like Harmony)
  • Fix "Cannot read properties of undefined (reading 'properties')" for certain tools-containing requests to /v1/chat/completions
  • Fix Error: EPERM: operation not permitted, unlink when auto-updating harmony

Build 1

  • Bug fixes resulting in significant improvements for in chat tool calling reliability
  • Fixed a bug where some old conversations won't load in the app
  • Fixed a bug where tool call will fail sometimes when used via OpenAI compatible API in non-streaming mode
  • Fix models not outputting thinking tags in v1/chat/completions
    • For gpt-oss:
      • message.content will not include reasoning content or special tags
      • This matches the behavior of o3-mini.
      • Reasoning content will be in choices.message.reasoning (stream=false) and choices.delta.reasoning (stream=true)
  • Fix "TypeError: Invalid Version" causing app functionality issues on machines with AMD+NVIDIA GPUs
  • Fix bug where MCP plugin chip name was not rendering in User Mode
  • Fix bug where search results were refreshing on click

Resources