Skip to content
Nx IDEDocs

AI Providers & Models4 min read

LM Studio (Local)

Local models with an OpenAI-compatible server

LM Studio is a desktop app for discovering, downloading, and running open-weight models locally — with a friendly GUI for picking quantizations and monitoring resource use. It exposes the models through a local OpenAI-style server that Nexgile Code connects to directly, so everything runs offline on your hardware.

Set up LM Studio

Section titled “Set up LM Studio”
  1. Install LM Studio from lmstudio.ai.
  2. Use its built-in browser to download a model. Favor instruction-tuned models with tool-calling support and pick the largest quantization your RAM/VRAM handles comfortably.
  3. Start the local server — this is the step people miss. Open LM Studio’s server (Developer) tab, load your model, and start the server. By default it listens on http://localhost:1234.

The product’s own settings panel reminds you of this: you must start LM Studio’s local server feature to use it with the agent.

Configure in Nx IDE

Section titled “Configure in Nx IDE”
  1. Open Settings → Providers and choose LM Studio.
  2. Base URL (optional) — leave empty for the default http://localhost:1234; set it only for a non-standard port or a server on another machine.
  3. Model ID — the picker lists the models your LM Studio server currently exposes, e.g. meta-llama-3.1-8b-instruct. If nothing shows up, the server isn’t running or has no model loaded.

Speculative decoding (optional)

Section titled “Speculative decoding (optional)”

LM Studio supports speculative decoding: a small “draft” model proposes tokens that the main model verifies, which can speed up generation noticeably on capable hardware.

  • Check Enable Speculative Decoding.
  • Pick a Draft Model ID — the panel warns that the draft model must be from the same model family as the main model for this to work correctly (e.g. a 1B Llama drafting for an 8B Llama).
  • Load the draft model in LM Studio too; if the panel reports no draft models found, confirm the server is running in server mode.

Picking models

Section titled “Picking models”
  • Coding-tuned first: models tuned for code and instruction following behave far better under the agent’s structured prompting than generic chat models.
  • Context window: configure a generous context length when loading the model in LM Studio — agent prompts are large, and a small window silently truncates instructions. Symptoms of a too-small window: the agent “forgets” what it was doing or emits malformed tool calls.
  • Bigger beats faster for agent reliability. If responses are poor, try a larger model or a less aggressive quantization before blaming the setup.

Behavior defaults

Section titled “Behavior defaults”

Nexgile Code applies a default temperature of 0 for LM Studio models — deterministic output suits code edits, and small local models drift badly at higher temperatures. Override it with Use custom temperature in the profile if you want more variety; see Model Behavior & Costs.

A working profile looks like this:

Provider:    LM Studio
Base URL:    (empty — uses http://localhost:1234)
Model ID:    meta-llama-3.1-8b-instruct
Speculative: enabled, draft lmstudio-community/llama-3.2-1b-instruct

Offline use

Section titled “Offline use”

Once the model files are downloaded, LM Studio needs no network at all. Combined with Nexgile Code the whole agent loop — prompts, code, diffs, terminal output — stays on your machine. The same realistic-expectations note from the settings panel applies here: the agent works best with highly capable models, and small local models may not handle complex multi-step tasks as expected. Keep tasks scoped, keep approvals on, and lean on checkpoints.

Choosing between the two local options: Ollama is CLI-first and scripts cleanly; LM Studio gives you a GUI, easy quantization switching, and speculative decoding. Nexgile Code treats both as first-class — run whichever fits your workflow, or both as separate configuration profiles.

Troubleshooting

Section titled “Troubleshooting”
  • Model picker is empty: the local server isn’t running, or no model is loaded into it. Start the server in LM Studio’s server tab and load a model, then reopen the settings panel.
  • “No draft models found”: the panel shows this when speculative decoding is enabled but the server exposes no candidate drafts — confirm LM Studio is running with server mode enabled and the draft model is downloaded.
  • Selected model flagged as unavailable: the panel warns when your saved Model ID no longer matches anything the server reports. Re-select from the picker.
  • Requests time out on big prompts: raise nexgile-code.apiRequestTimeout, and consider a smaller model or shorter context if generation is too slow for your hardware.
  • Garbled or non-working tool calls: the model is too small or not instruction-tuned for structured agent output; switch models rather than fighting the prompt.
Section titled “Related”