Under LM Studio are two different inference engines, depending on the machine. On Windows and Linux it uses llama.cpp to process GGUF model files. On Apple Silicon it uses MLX, Apple’s own framework, which processes vision input more quickly than llama.cpp can on the same Macs. There is no option to choose between them. The app picks.
Model discovery runs against Hugging Face’s catalogue, which can be queried from the left sidebar. Type Qwen or Gemma or DeepSeek and the results come back as variants rather than one file, each with its quantization level, size on disk and an estimate of the VRAM it will need. You read that estimate before you commit bandwidth to a 40 GB download. Downloads stop, resume and queue.
When loading a model, LM Studio automatically identifies the GPU and enables layer offloading: CUDA on NVIDIA, Vulkan on AMD, Metal on Apple Silicon, and AVX2 or AVX-512 instructions on the CPU where available. During model execution, the interface displays tokens per second, memory consumption and GPU load.
Developer Mode exposes settings the default view does not show: context length overrides, rope frequency, batch size, and per-layer GPU offloading instead of the automatic split.
Downloaded models are stored in a My Models tab, and chat history remains associated with the model that generated it. Conversations file into folders.
The Developer tab makes the app a server. Turn it on and a model responds via HTTP on localhost, or on the local network if you open it up. The endpoints are OpenAI’s and Anthropic’s, so tools written against either point to your machine rather than a vendor. There is a REST API of LM Studio’s own, as well as lmstudio-js and lmstudio-python SDKs. There is also a command-line utility, lms, that starts the same server from a terminal with lms server start, and a headless daemon, llmster, that runs the entire thing without the desktop app for servers and cloud instances.
Once that is running, the model does not have to be on the machine you are sitting at.
LM Link connects devices running LM Studio or llmster, and remote models are listed in the model loader next to the local ones. You load from a GPU rig in another room, an office workstation, or a cloud VM, as if the weights were on your own disk. The connection is end-to-end encrypted and based on custom Tailscale mesh VPNs. It is in preview, and is being rolled out in stages.
MCP servers make a model accessible to tools and data sources outside the app. One ships by default: a sandbox to run the model’s JavaScript or TypeScript using Deno. Adding others is manual work. There is no directory to browse and no install button — you edit an mcp.json file that describes the integration and provide the code yourself.
Bionic is the agent layer, targeting work and code. It generates documents, slides, PDFs and software, based on a locally running model or a frontier open model. Voice and audio input is processed on the device in multiple languages, and is not sent elsewhere.
