Skip to content

Set up a more robust local AI server (LM Studio, Ollama)

The bundled Understand Local model needs no setup and is the right choice for most people. Point Understand at LM Studio or Ollama instead when you want more than the bundled runtime gives you:

  • Bigger or different models — anything those runtimes can load, managed in their UI/CLI. Bigger generally means better answers but slower responses; it's a real tradeoff, not a free upgrade, and only pays off if the hardware underneath can actually run the larger model at a usable speed.
  • Better use of your hardware — their own GPU backends, quantizations, and tuning knobs.
  • One shared server — run the model on a beefy machine and let the whole team use it over your own network.

Both are still local: inference happens on your machine or a machine you control, and no API key is involved. (For the privacy picture see What can the AI access?; for a no-network deployment see Local AI with no cloud calls.)

Or share the bundled server

You don't have to switch runtimes to share one machine's model — Understand's own undaiserver can also be run for a whole team, and it powers batch generation from the command line via und ai. Use LM Studio or Ollama when you specifically want their model libraries or tuning.

In all cases the Understand side is the same: Data preferences → the ... button next to the AI Provider row, pick the provider, and confirm. You'll need to tick the data acknowledgement before OK is enabled — it applies to every provider other than the bundled one, even though these two never leave your network.

LM Studio

  1. Install LM Studio and enable Developer Mode (Settings → Developer).
  2. Download a model in LM Studio and load it.
  3. Start the local server from the Developer pane — the status should read Running. The default address is 127.0.0.1:1234.
  4. In Understand's provider dialog choose LM Studio, verify the address and port, pick the model, tick the acknowledgement, and click OK.

With LM Studio 0.4 or newer, Understand lists every model LM Studio has — loaded or not — and by default follows whichever model is currently loaded. On older versions it falls back to the OpenAI-compatible listing, which only shows loaded models.

If LM Studio is detected but no model appears, make sure a model is actually loaded (or that LM Studio's just-in-time loading is enabled) — without JIT, the fallback listing is empty until you load one.

Ollama

  1. Install Ollama and pull a model, e.g.:
ollama pull qwen3:4b
  1. Make sure the server is running (the Ollama app runs it automatically; otherwise run ollama serve). The default address is 127.0.0.1:11434.
  2. In Understand's provider dialog choose Ollama, verify the address and port, pick the model, tick the acknowledgement, and click OK.

Understand talks to Ollama's OpenAI-compatible endpoint, so any model ollama list shows is selectable.

Serve a whole team from one machine

Both runtimes can listen on the network instead of just 127.0.0.1:

  • LM Studio — in the Developer pane's server settings, enable serving on the local network.
  • Ollama — start the server with OLLAMA_HOST=0.0.0.0 in its environment.

Then, in each user's Understand provider dialog, replace 127.0.0.1 with the server's hostname or IP. The default port for the chosen provider is filled in automatically if you leave it off, and plain http is fine — HTTPS is only enforced when an API key is configured.

Only point at machines you control

Everything on this page keeps prompts — including code context — inside your own network as long as the endpoint is yours. Entering a hosted third-party URL here makes it a cloud provider in all but name. See Local AI with no cloud calls.