Setup LM Studio API Server: OpenAI-Compatible Local Endpoint 2026

Problem: LM Studio Won't Serve API Requests to Your Code

The LM Studio API server lets you swap any OpenAI API call for a local LLM endpoint — zero cost, no rate limits, no data leaving your machine. But the server tab is easy to misconfigure, and the default OpenAI SDK setup points at api.openai.com, not localhost:1234.

You'll learn:

• How to start the LM Studio local server and load a model correctly
• How to point the OpenAI Python and Node.js SDKs at localhost:1234
• How to verify requests end-to-end with curl and confirm streaming works

Time: 20 min | Difficulty: Intermediate

Why the LM Studio API Server Exists

LM Studio ships a built-in HTTP server that mimics the OpenAI REST API. Any client that speaks OpenAI — LangChain, LlamaIndex, Continue.dev, your own scripts — can talk to it by changing one URL and removing the API key check.

The server exposes three endpoints:

No authentication is required on localhost. For LAN access you'll bind to 0.0.0.0 and optionally add a static API key — covered in Step 4.

Request flow: your code sends an OpenAI-format payload → LM Studio routes it to the loaded GGUF model → streams tokens back over HTTP

Step 1: Download LM Studio and Load a Model

Download LM Studio 0.3.x from lmstudio.ai — available for macOS (Apple Silicon + Intel), Windows, and Ubuntu 22.04+. The current stable release is 0.3.6.

Open LM Studio, go to the Search tab, and pull a model. For this guide we use Mistral 7B Instruct v0.3 Q4_K_M — it fits in 8 GB RAM and performs well for chat.

Model: mistralai/Mistral-7B-Instruct-v0.3
Quant: Q4_K_M  (4.37 GB)
Min RAM: 8 GB

Wait for the download to finish (the progress bar fills green). You'll see the model in My Models.

Step 2: Start the Local Server

Click the Local Server tab (the <-> icon in the left sidebar).

Configure these fields before clicking Start:

Click Start Server. You should see:

[LM Studio] Server listening on http://localhost:1234
[LM Studio] Model loaded: mistral-7b-instruct-v0.3.Q4_K_M.gguf

If the server starts but shows "No model loaded": Go back to the model selector dropdown in the Server tab and explicitly choose your model. The chat model loaded in the Chat tab does NOT carry over automatically.

Step 3: Verify with curl

Before touching any SDK, confirm the raw endpoint works. Open a terminal:

# Check that the server lists your model
curl http://localhost:1234/v1/models | python3 -m json.tool

Expected output:

{
  "data": [
    {
      "id": "mistral-7b-instruct-v0.3",
      "object": "model"
    }
  ]
}

Now send a chat completion:

curl http://localhost:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mistral-7b-instruct-v0.3",
    "messages": [{"role": "user", "content": "Reply with: API is working"}],
    "temperature": 0.1,
    "max_tokens": 20
  }'

You should see: A JSON response with "content": "API is working" inside choices[0].message.

If you get Connection refused: The server is not running. Return to Step 2 and confirm the green "Running" indicator is visible in LM Studio.

If you get 500 Internal Server Error: The model ID in "model" doesn't match — copy the exact id string from the /v1/models response.

Step 4: Connect the OpenAI Python SDK

Install the SDK if you haven't already:

pip install openai

Change only base_url and api_key. Everything else — ChatCompletion.create, streaming, system messages — works identically to the real OpenAI API:

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:1234/v1",  # LM Studio server
    api_key="lm-studio",                  # any non-empty string — LM Studio ignores it
)

response = client.chat.completions.create(
    model="mistral-7b-instruct-v0.3",     # must match /v1/models id exactly
    messages=[
        {"role": "system", "content": "You are a concise assistant."},
        {"role": "user", "content": "What is the OpenAI-compatible endpoint LM Studio exposes?"},
    ],
    temperature=0.3,
    max_tokens=256,
)

print(response.choices[0].message.content)

Streaming version — same change, just add stream=True:

stream = client.chat.completions.create(
    model="mistral-7b-instruct-v0.3",
    messages=[{"role": "user", "content": "Count to 5 slowly."}],
    stream=True,  # tokens arrive as they're generated
)

for chunk in stream:
    delta = chunk.choices[0].delta.content
    if delta:
        print(delta, end="", flush=True)

Expected output: Tokens print one by one instead of waiting for the full response.

Step 5: Connect with Node.js / TypeScript

npm install openai

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "http://localhost:1234/v1",
  apiKey: "lm-studio", // required by SDK constructor — value is ignored by LM Studio
});

async function main() {
  const response = await client.chat.completions.create({
    model: "mistral-7b-instruct-v0.3",
    messages: [{ role: "user", content: "Explain GGUF in one sentence." }],
    temperature: 0.2,
    max_tokens: 80,
  });

  console.log(response.choices[0].message.content);
}

main();

Run with:

npx tsx index.ts
# or: node --experimental-strip-types index.ts  (Node 22+)

Step 6: LAN Access (Optional)

To reach the server from other machines on your network — useful for shared dev environments:

1. In LM Studio Server tab, set Bind to → 0.0.0.0
2. Restart the server
3. Find your local IP: ipconfig (Windows) or ifconfig | grep inet (macOS/Linux)
4. Update base_url in your client: http://192.168.x.x:1234/v1

For basic auth, LM Studio 0.3.x lets you set a static API Key in the server settings. Set it, then pass it as api_key in the SDK — LM Studio will reject requests without it.

Windows firewall note: You'll need to allow inbound TCP on port 1234. Run in an elevated PowerShell:

New-NetFirewallRule -DisplayName "LM Studio API" -Direction Inbound -Protocol TCP -LocalPort 1234 -Action Allow

LM Studio vs Ollama: API Server Comparison

Choose LM Studio if: you want a GUI to browse, download, and test models before wiring them to code.

Choose Ollama if: you need Docker, multiple concurrent models, or a headless server on a remote machine.

Verification

Run the full stack check:

# 1. Server status
curl -s http://localhost:1234/v1/models | python3 -c "import sys,json; d=json.load(sys.stdin); print('Model ID:', d['data'][0]['id'])"

# 2. Round-trip latency
time curl -s http://localhost:1234/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"mistral-7b-instruct-v0.3","messages":[{"role":"user","content":"hi"}],"max_tokens":5}' \
  > /dev/null

You should see:

• Model ID: mistral-7b-instruct-v0.3
• A real time under 5 seconds for first-token latency on CPU, under 1s with GPU offloading

What You Learned

• LM Studio's local server is a drop-in OpenAI API replacement — change base_url and a dummy api_key, nothing else
• The model in the Server tab must be set explicitly; the Chat tab model does not carry over
• Streaming, system messages, and temperature work identically to the OpenAI API
• For multi-model or Docker deployments, Ollama is the better fit

Tested on LM Studio 0.3.6, Mistral 7B Instruct v0.3 Q4_K_M, Python 3.12, openai SDK 1.30, Node.js 22, macOS 15 (M2) & Windows 11

FAQ

Q: Do I need an OpenAI API key to use LM Studio's server? A: No. LM Studio ignores the api_key field on localhost. Pass any non-empty string like "lm-studio" to satisfy the SDK constructor.

Q: Why does /v1/chat/completions return a 500 error? A: The model field in your request doesn't match the ID returned by /v1/models. Copy the exact string from GET /v1/models — it usually includes the quantization suffix.

Q: What is the minimum RAM to run a model with the LM Studio API server? A: 8 GB RAM for a Q4_K_M 7B model with CPU inference. For GPU offloading you need at least 6 GB VRAM to fit most layers of a 7B model. Add 2–3 GB headroom for the OS.

Q: Can LM Studio serve multiple models at the same time? A: No — LM Studio 0.3.x loads one model per server instance. For concurrent models use Ollama or run two separate LM Studio instances on different ports.

Q: Does LM Studio's API support function calling / tool use? A: Yes, for models that have been fine-tuned for tool use (e.g., Mistral 7B Instruct v0.3, Llama 3.1 Instruct). Pass a tools array the same way you would with the OpenAI SDK — LM Studio forwards the schema to the model.