LM Studio runs llama.cpp (GGUF) or MLX models locally, as a GUI app or the headless llmster daemon. For install and product docs, see lmstudio.ai.

Quick start

Install and start the server

Install LM Studio (desktop) or `llmster` (headless), then start the server:

```bash
lms server start --port 1234
```

Or run the headless daemon:

```bash
lms daemon up
```

If using the desktop app, enable JIT for smooth model loading; see the
[LM Studio JIT and TTL guide](https://lmstudio.ai/docs/developer/core/ttl-and-auto-evict).

Set an API key if auth is enabled

```bash
export LM_API_TOKEN="your-lm-studio-api-token"
```

If LM Studio authentication is disabled, leave the API key blank during setup. See
[LM Studio Authentication](https://lmstudio.ai/docs/developer/core/authentication).

Run onboarding

```bash
apus onboard
```

Choose `LM Studio`, then pick a model at the `Default model` prompt.

The server URL prompt also accepts host shorthand such as `localhost:1234`.
Invalid URLs stay in the prompt so you can correct them before model discovery.

On a fresh guided setup, Apus first queries `/api/v1/models` on the
default or configured LM Studio host. An existing LLM is offered automatically
only when LM Studio reports tool training and at least 16K of effective
context. For loaded models, the loaded instance context takes precedence over
the larger advertised maximum. The same CLI/macOS setup ladder verifies the
route with a real completion before saving it. The automatic check never
downloads a model and ignores embedding-only catalog entries.

Change the default model later:

apus models set lmstudio/qwen/qwen3.5-9b

LM Studio model keys use an author/model-name format (e.g. qwen/qwen3.5-9b); Apus model refs prepend the provider: lmstudio/qwen/qwen3.5-9b. Find the exact key for a model by running the command below and looking at the key field:

curl http://localhost:1234/api/v1/models

After installing another model in LM Studio, refresh Apus's model list:

apus models list --provider lmstudio --refresh

In the default merge mode, refresh adds discovered models while preserving your configured rows and their authored metadata, including names and context limits. An empty configured models array also supports discovery on an unauthenticated server. models.mode: "replace" keeps only explicitly configured models.

Non-interactive onboarding

apus onboard --non-interactive --accept-risk --skip-health --auth-choice lmstudio

Or specify base URL, model, and API key explicitly:

apus onboard \
  --non-interactive \
  --accept-risk \
  --skip-health \
  --auth-choice lmstudio \
  --custom-base-url http://localhost:1234/v1 \
  --lmstudio-api-key "$LM_API_TOKEN" \
  --custom-model-id qwen/qwen3.5-9b

--custom-model-id takes the model key as returned by LM Studio (e.g. qwen/qwen3.5-9b), without the lmstudio/ provider prefix. Pass --lmstudio-api-key (or set LM_API_TOKEN) for authenticated servers; omit it for unauthenticated servers and Apus stores a local non-secret marker instead. --custom-api-key is still accepted for compatibility, but --lmstudio-api-key is preferred.

This writes models.providers.lmstudio and sets the default model to lmstudio/<custom-model-id>. Providing an API key also writes the lmstudio:default auth profile.

Add --json for a machine-readable result. Connection, HTTP, and model-selection failures return a JSON error with the same recovery guidance as human output and exit nonzero without applying the proposed provider configuration.

Interactive setup can additionally prompt for a preferred load context length and applies it across the discovered models it saves to config.

Configuration

Streaming usage compatibility

LM Studio doesn't always emit an OpenAI-shaped usage object on streamed responses. Apus recovers token counts from llama.cpp-style timings.prompt_n / timings.predicted_n metadata instead. Any OpenAI-compatible endpoint resolved as a local endpoint (loopback host) gets this same fallback, which covers other local backends such as vLLM, SGLang, llama.cpp, LocalAI, Jan, TabbyAPI, and text-generation-webui.

Gemma 4 tool-call recovery

For Gemma 4 models using openai-completions, Apus recovers complete standalone <|tool_call>call:...<tool_call|> batches if the server returns them as text, including when the call is the last content before finish_reason: "stop". Recovery preserves raw string arguments and requires complete argument objects; incomplete calls, prose, and code examples remain text. Truncated, filtered, cancelled, or unterminated streams do not authorize recovered calls. If native tool calls also appear in the stream, they remain authoritative and raw text is not promoted.

Thinking compatibility

When LM Studio's /api/v1/models discovery reports model-specific reasoning options, Apus exposes matching reasoning_effort values (none, minimal, low, medium, high, xhigh) in model compat metadata. Some LM Studio builds advertise a binary UI option (allowed_options: ["off", "on"]) while rejecting those literal values on /v1/chat/completions; Apus normalizes that binary shape to the six-level scale before sending requests, including for older saved config that still has off/on reasoning maps.

For graded options, fresh discovery maps max (and Ultra's provider effort) to the highest advertised canonical effort, regardless of option order. For example, off, low, medium maps max to medium. Existing saved graded reasoningEffortMap values remain explicit configuration; rerun LM Studio setup to regenerate them from current server metadata.

Explicit configuration

{
  models: {
    providers: {
      lmstudio: {
        baseUrl: "http://localhost:1234/v1",
        apiKey: "${LM_API_TOKEN}",
        api: "openai-completions",
        models: [
          {
            id: "qwen/qwen3-coder-next",
            name: "Qwen 3 Coder Next",
            reasoning: false,
            input: ["text"],
            cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
            contextWindow: 128000,
            maxTokens: 8192,
          },
        ],
      },
    },
  },
}

Model instances and context

With preload enabled, Apus routes chat requests to a loaded instance with enough context for the selected model budget. A newly loaded instance is addressed by the identifier returned by LM Studio. Your configured model reference and conversation model identity keep the canonical model key.

With preload enabled, embedding requests also check that their model is loaded and route to the instance prepared for the configured context length. This avoids truncating input through a smaller loaded instance and lets memory embeddings recover after model eviction even when LM Studio JIT loading is disabled. Embedding model and cache identity keep the canonical model key.

Disabling preload

LM Studio supports just-in-time (JIT) model loading, loading models on first request. Apus preloads models through LM Studio's native load endpoint by default, which helps when JIT is disabled. To let LM Studio's JIT, idle TTL, and auto-evict behavior own model lifecycle instead, disable Apus's preload step:

{
  models: {
    providers: {
      lmstudio: {
        baseUrl: "http://localhost:1234/v1",
        api: "openai-completions",
        params: { preload: false },
        models: [{ id: "qwen/qwen3.5-9b" }],
      },
    },
  },
}

LAN or tailnet host

Use the LM Studio host's reachable address, keep /v1, and make sure LM Studio is bound beyond loopback on that machine:

{
  models: {
    providers: {
      lmstudio: {
        baseUrl: "http://gpu-box.local:1234/v1",
        apiKey: "lmstudio",
        api: "openai-completions",
        models: [{ id: "qwen/qwen3.5-9b" }],
      },
    },
  },
}

lmstudio automatically trusts its configured endpoint for model requests, including loopback, LAN, and tailnet hosts (except metadata, link-local, and local-use NAT64 64:ff9b:1::/48 origins). Any custom/local OpenAI-compatible provider entry gets the same exact-origin trust. Requests to a different private host or port still require models.providers.<id>.request.allowPrivateNetwork: true; set it to false to opt out of the default trust.

Troubleshooting

Context overflow

When LM Studio rejects a prompt because it exceeds the loaded context, Apus attempts compaction and retry, including for older servers that report Trying to keep the first N tokens and the model is loaded with context length of only M tokens, which is not enough. If a single message is too large, shorten it or load the model with a larger context; compacting earlier history cannot shrink that message. Server-side truncation without an overflow error does not trigger this recovery.

Model discovery failures

When a configured server cannot list models, Apus reports an unavailable catalog or a catalog authentication rejection. A refresh can keep the last successful inventory when the connection and credentials still match. A successful empty response clears discovered models; explicitly configured models remain available without discovery. Restore the server connection or correct its credentials, then refresh the model list.

LM Studio not detected

Make sure LM Studio is running:

lms server start --port 1234

If authentication is enabled, also set LM_API_TOKEN. Verify the API is reachable:

curl http://localhost:1234/api/v1/models

Authentication errors (HTTP 401)

Related