You decide to run open-weight models on your own machine, and the same two names show up almost immediately: Ollama and LM Studio. Both sit on Georgi Gerganov’s llama.cpp and both keep the weights on your own GPU and RAM. What changes is how you work with them once they’re installed. One leaves a server running behind the scenes, waiting for a request even when you close the window. The other opens a window where you pick a model and start chatting. If you want autocomplete in your editor, the first one is probably your answer. If you want to compare how different quantization levels answer the same prompt, it’s the second.
A terminal-first tool and a window-first tool
Ollama has a desktop app too. Open it and follow the setup prompts to sign in for cloud models or pick a local model. On macOS there’s also an Apps menu that connects the Claude and ChatGPT desktop apps to your Ollama models. With ChatGPT, that only covers Codex mode: regular Chat and voice keep your usual ChatGPT connection. The center of gravity is still the CLI and the HTTP API. Install it and it binds to 127.0.0.1 on port 11434 by default. Run ollama run gemma4 and it downloads the model and starts a chat on your machine. If you installed the app or the Linux systemd service, that server keeps running in the background, so an editor extension or a script can still call it after you close the terminal.
LM Studio starts from the desktop app instead. You search and download models from Hugging Face inside the app, then chat with them right there. The same app also attaches documents to a chat and answers questions about them entirely offline (RAG), and installs MCP servers to use with local models. The lms CLI and the headless llmster daemon live in the developer docs.
The engine is the same on both sides. LM Studio also offers Apple’s MLX runtime on Apple Silicon Macs. No doc page pits the two runtimes against each other on speed, so running the same model through each is the reliable way to tell.
Ports and endpoints matter more than the numbers
Local models rarely stay inside a chat window. Most people wire them into an editor extension, an agent framework, or a script, and that’s where the HTTP surface decides things.
Ollama’s local API splits into /api and /v1. Its own endpoints include /api/chat for conversations and /api/generate for single prompts. The /v1 side is the OpenAI-compatible layer, serving /v1/chat/completions and /v1/responses. Anthropic clients can connect too: set the base URL to http://localhost:11434 and the client appends /v1/messages. If you already have a script written against the Claude Messages format, change the base URL and put a local model name in model. Ollama supports a subset of the Messages API, so fields like tool_choice and metadata are missing.
For the ecosystem side, the clearest source is the Community Integrations list in the Ollama README. Open WebUI, Dify, Cline, Continue, and LangChain all appear there. That just means Ollama’s README lists them. It doesn’t tell you how prominently each one features it, and that varies by tool.
LM Studio’s server comes up when you toggle the Start server switch in the Developer tab. The docs use port 1234 in their examples, as in lms server start --port 1234 and a curl to http://localhost:1234/.... The lms server start reference says that with --port left out it uses the last used port, and the OpenAI compatibility page notes that its examples assume 1234. So don’t treat 1234 as guaranteed. Check the port your server actually started on.
Its API covers OpenAI-compatible /v1/chat/completions and /v1/responses, Anthropic-compatible /v1/messages, and its own v1 REST API at /api/v1/chat. The native API shipped officially in 0.4.0 and covers stateful chats, MCP calls over HTTP, and model load and unload. For a server or CI box with no GUI, there’s the llmster daemon and the lms CLI.
When you want an always-on coding endpoint
If you need autocomplete or a background endpoint for scripts, Ollama asks less of you. You don’t have to keep an app window open while your editor runs. Memory is handled by the server: by default a model stays in memory for 5 minutes and then unloads, and the keep_alive parameter or the OLLAMA_KEEP_ALIVE environment variable changes that window. A negative value keeps it loaded, and 0 unloads it right after the response.
A Modelfile lets you bundle a system prompt, temperature, and context length into one custom model name, which is handy when each project needs its own instructions.
When you’re comparing prompts and documents
For eyeballing different models and settings, LM Studio is the easier place to work. You can search Hugging Face from inside the app and download a specific quantization, and the offline document chat runs in the GUI.
Memory is where the two tools look more alike than they’re usually described. Both can unload idle models. Ollama does it after 5 minutes by default. LM Studio uses just-in-time loading, with a default TTL of 60 minutes for models loaded that way, and its Developer tab server settings include switches to unload unused JIT models automatically. Models loaded by hand with lms load have no TTL and stay until you unload them. Knowing the default number matters more than the claim that either one “handles it for you.”
Running both on one machine
On the example ports, 11434 and 1234, the two servers don’t collide, so installing both is fine. Keep your coding extension pointed at Ollama and open LM Studio only when you want to try a new model. Model files, though, pile up separately. Ollama stores them at ~/.ollama/models on macOS, /usr/share/ollama/.ollama/models on Linux, and C:\Users\%username%\.ollama\models on Windows, and OLLAMA_MODELS moves that directory. LM Studio creates a publisher/model folder tree under ~/.lmstudio/models/. Pull the same 30B-class model on both and you’re paying for the disk twice.
The fastest way to decide is to check what you already have. If your VS Code setup has the Ollama extension, or your Cline or Continue config already points at localhost:11434, you’ve made the choice already. If nothing is set up, download one model and finish a single conversation today. That run will tell you more than another comparison table.