Pinned
We made @lmstudio's MLX Engine a lot faster in the latest release.
Read the technical deep dive from @ostensiblyneil.
P.S. it's all open source!
LM Studio Developers
106 posts
LM Studio Developers
@lmstudiodevs
Joined June 2025
- |￣￣￣￣￣￣￣￣￣￣￣￣￣￣| | lms get openai/gpt-oss-20b | |＿＿＿＿＿＿＿＿＿＿＿＿＿＿| (\__/) || (•ㅅ•) || / づOur open models are here. Both of them. openai.com/open-models
- 🧪 LM Studio 0.3.21 build 2 (beta) with bug fixes for Qwen3 Coder tool calling is now available. Update in-app or head to beta releases on the website.
- In LM Studio 0.3.23, you can enable a setting similar to llama.cpp's `--cpu-moe`. If you're low on VRAM and need to split the model between GPU and CPU, give this setting a try.
LM Studio 0.3.23 is out now! This release contains lots of reliability fixes, and an advanced setting to make MoE models faster on low VRAM machines. - If you're using gpt-oss via LM Studio's /v1/chat/completions API, try to turn this setting on to separate reasoning from regular content. Need to update to v0.3.22 first!
- LM Studio 0.3.22 b2 ships with Harmony, OpenAI's prompt formatting library for gpt-oss. LM Studio uses Harmony to construct the input to gpt-oss models, for both the llama.cpp as well as the MLX engine. Both through the app's chat UI and API.
- New in LM Studio 0.3.24b1 (beta): Global setting to always use the max context length supported by the model.
- Replying to @lmstudiodevsWe ship it as an updatable "Framework", which is an extension to the LM Studio runtime. See more details: > Mac: Cmd + Shift + R > PC: Ctrl + Shift + R
- Replying to @lmstudiodevsSee the llama.cpp PR that added this feature here:
