Ollama vs LM Studio vs Jan: which local AI app should you use?

The quick answer

Ollama, LM Studio and Jan all download open models such as Llama, Gemma, Qwen, gpt-oss and DeepSeek-R1 and run them on your own PC. All three build on the open-source llama.cpp engine, so the real differences are in how you use them, their licences and what else they connect to. ToolsBharat hasn't benchmarked them, and none of the three vendors publishes speed comparisons, so this guide compares only what their own docs say.

Side by side

Licence and source code

• Ollama: open source, MIT licence.
• LM Studio: proprietary. Its terms grant a licence to use the app "solely for Your personal and / or internal business purposes" and mention that certain features may be offered for a fee. The lms command-line tool that ships with it is MIT-licensed.
• Jan: open source, Apache 2.0 licence.

How you use it

• Ollama: runs in the background with a tray icon. You mostly work through the ollama command (for example ollama run gemma4) or through other apps that connect to it. Its ollama launch command sets up coding tools such as Claude Code, Codex and OpenCode to use Ollama models.
• LM Studio: a full desktop app. The Discover tab searches Hugging Face, the Chat tab is where you talk to the model, and you can attach documents to chat with them offline. It can also connect MCP servers.
• Jan: a ChatGPT-style desktop app. It runs local models from Hugging Face and can also connect to cloud providers such as OpenAI, Anthropic, Mistral and Groq if you add your own keys. It supports custom assistants and MCP.

Windows requirements (from each vendor)

• Ollama: Windows 10 22H2 or newer. NVIDIA driver 551.61+ for NVIDIA GPUs. AMD Radeon through ROCm v7 or Vulkan. At least 4 GB of disk for the app.
• LM Studio: x64 or ARM (Snapdragon X Elite). AVX2 is required on x64. At least 16 GB of RAM recommended, and at least 4 GB of dedicated VRAM recommended.
• Jan: Windows 10 or higher. x64 with AVX2, or ARM64. 8 GB of RAM minimum (16 GB recommended). 6 GB of VRAM minimum for NVIDIA, AMD or Intel Arc GPUs. 10 GB of free space minimum.

Local API for other apps

• Ollama: http://localhost:11434, plus OpenAI- and Anthropic-compatible endpoints.
• LM Studio: OpenAI-compatible endpoints on localhost or your local network, and a headless mode (llmster) for machines without a screen.
• Jan: an OpenAI-compatible server at localhost:1337.

Privacy, in the vendors' words

• Ollama: "We don't see your prompts or data when you run locally." Optional cloud models run on Ollama's servers. You can switch cloud features off with OLLAMA_NO_CLOUD=1.
• LM Studio: "Nothing you enter into LM Studio when chatting with LLMs leaves your device." Searching for models, downloading them and checking for app updates do use the internet.
• Jan: its Windows install page says Jan keeps all data locally and "Nothing is sent to the cloud". Data lives in AppData\Roaming\Jan\data. If you connect a cloud provider, those chats go to that provider.

How to choose, step by step

1. Check your processor. If your x64 CPU lacks AVX2, LM Studio and Jan list it as required, so Ollama is the one to try.
2. Check your memory. With 8 GB of RAM and no graphics card, any of the three will only handle small models. LM Studio's own recommendation is 16 GB.
3. Decide how you like to work. Prefer clicking? Choose LM Studio or Jan. Prefer typing commands, or want to plug a local model into Claude Code, Codex or VS Code? Choose Ollama.
4. Decide whether open source matters to you. If it does, choose Ollama or Jan. LM Studio is closed source.
5. Do you also want cloud models in the same window? Jan connects to several cloud providers. Ollama offers its own cloud models.
6. Install just one to start. Each app keeps its own model folder, so installing all three and downloading the same model three times wastes disk space.

Troubleshooting when switching between them

• Port already in use: if two apps run local servers at the same time, close the one you're not using. Ollama uses port 11434 and Jan uses 1337.
• Disk filling up: delete models you no longer use in each app. In Ollama that's ollama rm <model>. In Jan, models live in its data folder.
• The model is slow in every app: that's your hardware, not the app. Try a smaller model or a lower quantization, such as a 4-bit option.

Ready to start? See how to install Ollama on Windows or how to run DeepSeek-R1 locally.