If you want to run a chatbot on your own PC rather than in the cloud, Jan and LM Studio are two of the easiest ways to start. Both are free desktop apps for Windows, macOS and Linux. Both let you download open models, chat with them offline and run a local server that other apps can use. The main difference is philosophy: Jan is open source, while LM Studio is closed-source software with its own terms of use.
Forge PC has not benchmarked Jan against LM Studio, so this comparison does not make speed claims. Both apps use llama.cpp as an engine for GGUF models on Windows, so performance on the same hardware and model is likely to depend more on settings than on the app. Everything below comes from each project's official documentation, licences and release notes, checked on 14 November 2026.
Quick comparison
| Jan | LM Studio |
|---|
Licence | Open source, Apache 2.0 | Closed source; free licence for personal and internal business use |
Latest version (14 November 2026) | v0.8.8, released 28 October 2026 | 0.4.28, released 31 October 2026 |
Engines | Bundled llama.cpp; MLX on Apple Silicon | llama.cpp and MLX |
NVIDIA GPUs | Bundled CUDA 13 engine | Supported |
AMD and Intel Arc GPUs | Bundled Vulkan backend | Supported through its runtimes (check for your card) |
Local API server | OpenAI-compatible, at 127.0.0.1:1337 by default | OpenAI- and Anthropic-compatible, at localhost:1234 by default |
Minimum PC (Windows) | AVX2 CPU, 8GB RAM (16GB recommended), 6GB VRAM for GPU use | AVX2 CPU, 16GB RAM recommended, 4GB+ VRAM recommended |
Intel Macs | Bundled local engine not supported | Apple Silicon only |
Licence and openness
Jan is published on GitHub under the Apache 2.0 licence, which means you can read, modify and redistribute the code. The project had about 45,700 GitHub stars on 14 November 2026. If you care about being able to inspect what an app does with your data, that openness is Jan's biggest advantage.
LM Studio is free to download, but it is not open source. Its app terms grant "a non-exclusive, non-transferable license to use the Software solely for Your personal and / or internal business purposes", and describe its source code as "valuable trade secrets". For home users and most office use, that licence is fine, but you cannot inspect or modify the app itself. Recent versions also promote Bionic, LM Studio's separate agent app for work and coding.
Hardware support
Both apps run models on your graphics card when possible, and both rely on video memory. Our VRAM guide explains how much you need for different model sizes.
Jan documents its GPU support clearly. On Windows, it supports NVIDIA, AMD and Intel Arc graphics cards. NVIDIA cards use Jan's bundled CUDA 13 engine, with no need to install the CUDA Toolkit, and cards or drivers that cannot use CUDA 13 fall back to Vulkan. AMD and Intel Arc cards use the bundled Vulkan backend. On Macs, Apple Silicon is accelerated through Metal and MLX, but Jan says its bundled local engine does not support Intel Macs.
LM Studio recommends at least 4GB of dedicated VRAM and 16GB of system RAM on Windows, and requires an AVX2-capable processor. It runs models through llama.cpp and Apple's MLX, and on Macs it requires Apple Silicon. Its changelog shows ongoing work on AMD support, such as version 0.4.27 adding detection of the Radeon 8065S for its ROCm runtime.
If you have an AMD or Intel card, our guides to running local AI on a Radeon GPU and choosing NVIDIA vs AMD for local AI go further.
Finding and running models
LM Studio is known for its model browser, which helps you pick a model and quantisation that fits your hardware. Our guide to choosing models in LM Studio walks through it, and our quantisation explainer explains the file sizes.
Jan also downloads models from inside the app, and its documentation covers connecting cloud providers, with OpenAI and Anthropic among those named on its homepage. That lets you switch between local and cloud models in one interface. It also supports MCP (Model Context Protocol) tools.
Local API server
Both apps can act as a local server, so other tools, coding assistants and scripts can use your local model as if it were a cloud API.
Jan runs an OpenAI-compatible server at http://127.0.0.1:1337/v1 by default, with an optional API key. Version 0.8.8 added a Trusted Hosts setting, so web pages can only call the local server if their origin is trusted, and server-side MCP tool execution stays opt-in.
LM Studio serves models from its Developer tab, on localhost or across your network. It offers OpenAI-compatible endpoints at http://localhost:1234/v1 and Anthropic-compatible endpoints, plus its own REST API and client libraries, and the server can also be started with LM Studio's command-line tool, lms. Our LM Studio server guide covers the setup.
Which should you choose?
Choose Jan if you want open-source software you can audit, a single app that mixes local and cloud models, or straightforward Vulkan support for an AMD or Intel Arc card on Windows.
Choose LM Studio if you want the most polished model browser, Anthropic-compatible endpoints for coding tools, or a separate command-line tool (lms) for managing models and the server.
If you prefer a background service rather than a desktop app, see our Ollama vs LM Studio comparison. If you want a web interface for a household or small team, our Open WebUI vs AnythingLLM comparison may suit you better.
Our verdict
Jan and LM Studio are both free, capable ways to run AI models on your own PC, and both use llama.cpp under the hood for GGUF models. Choose Jan if open-source code, an Apache 2.0 licence and built-in Vulkan support for AMD and Intel Arc cards matter to you. Choose LM Studio if you want the most polished model browser and both OpenAI- and Anthropic-compatible local endpoints, and you are happy with closed-source software licensed for personal and internal business use. Forge PC has not benchmarked the two apps; this comparison relies on their official documentation, licences and release notes.