Skip to content
Search

LM Studio

lastmod 2026-04-08created 2024-05-145 edits

A tool that makes it easy to run LLMs locally

Features

  • Supports finding, downloading, and running LLMs locally and offline
    • Models such as LLaMa, Falcon, MPT, StarCoder, GPT-Neo-X
  • Mac/Windows/Linux (beta)
  • Models can be used through the built-in chat UI or an OpenAI-compatible local server
  • Supports ggml-based Llama, MPT, StarCoder models, etc. from HuggingFace
    • (e.g. Llama 2, Orca, Vicuna, Nous Hermes, WizardCoder, MPT, etc.).
  • Minimum specs and requirements
    • Apple Silicon Mac (M1/M2/M3) with macOS 13.6 or later
    • Windows / Linux PC (processor that supports AVX2)
    • 16GB+ RAM recommended. 6GB+ VRAM recommended on PC
    • Supports NVIDIA/AMD GPUs

LM Studio privacy and usage

  • LM Studio does not collect or monitor user data, and all data stays locally on the user's device
  • Free for personal use; for work use, you need to contact the team

GN⁺'s opinion

  • LM Studio is an attractive solution for privacy-conscious users, and the ability to run LLMs offline is the most important part.
  • It is interesting that user data is stored only locally and that a wide range of models is supported, and it is beneficial that it is free for individual users.
  • The hardware requirements are relatively high, making it suitable for users of recent devices, and it is positive that work use is being planned.

Hacker News comments

  • Shared a link showing that the same thing can be done with open-source code using Ollama and a UI.
  • Positive reviews of LM Studio and its features:
    • Local model runtime
    • Model catalog
    • UI for easily chatting with models
    • OpenAI-compatible API
    • Includes plugins such as RAG (using ChromaDB)
    • Well positioned to take advantage of new features in the OS ecosystem
    • Unfortunately, LM Studio itself is not open source
  • Wants to try uncensored models. Found the "TheBloke/Luna-AI-Llama2-Uncensored-GGML" model and downloaded only the first file to check whether it works. Wonders whether the remaining files also need to be downloaded, and whether the model is really uncensored on certain sensitive questions.
  • A beginner question about LLMs: what happens if you feed the "Lord of the Rings" series into an LLM, whether a basic "ignorant" LLM exists, and whether this could be used to create a Middle-earth expert.
  • Amusing qualifications for a senior engineering role: a deep understanding of computers and software, the ability to write algorithms and programs, etc. are required, but these requirements are hard to take seriously.
  • Shared a link introducing a simple UI they made personally as an attempt to simplify complex setups.
  • Curious about the potential user base of LLM software and wants to see information or speculation about market segmentation. Despite the complex setup, finds value in this interface and in chat and model management.
  • Suggestions for improving the app's UI:
    • The default state of the chat box field needs improvement
    • Unclear how to set CUDA acceleration before loading a model
    • Some HuggingFace models are not listed and no reason is given
    • Scrolling in the accordion parts of the interface responds only to the mouse wheel
    • The server tab is very useful for initial debugging
  • A question about problems that can arise when using various LLM models (Mistral, Llama 2, Code Llama, Orca Mini, Vicuna) (50% hallucination, recommending code for nonexistent APIs, reproducing outdated StackOverflow answers, etc.), and a request to share cases of actually using these models.
  • After installing and running LM Studio, CPU usage rose to 30% and laptop fan noise increased. After killing the LM Studio process and reopening it, background usage dropped to 5%.

See also

Favorite site

Articles

author Zer0 sitemap