What you will prove
- The model is stored and can be loaded on your machine.
- Local chat works without a paid inference API.
- The server responds on localhost:1234.
- Another application can use an OpenAI-compatible endpoint.
LM Studio is not the model
LM Studio is the application and runtime layer. Qwen, Gemma, Mistral, Llama, and other families are models it can load. Memory use depends on the model, format, quantization, context, caches, runtime, and the rest of your system.
LM Studio or Ollama?
LM Studio makes model discovery, loading, and configuration visible in a GUI. Ollama is especially convenient for terminal workflows and automation. You do not need a universal winner: testing a similar model in both helps separate model quality from runtime experience.
Check compatibility before downloading
- macOS: Apple Silicon M1-M4 and macOS 14 or newer. Intel Macs are currently unsupported.
- Windows: x64 with AVX2 or supported ARM systems. 16 GB RAM and 4 GB dedicated VRAM are current recommendations, not guarantees.
- Linux: x64 or ARM64; the desktop app is distributed as AppImage and documents Ubuntu 20.04 or newer.
Start with a small instruct model, roughly 1B-8B, and a modest 4K-8K context. The goal is to verify the stack, not to force the largest model into memory.
Install LM Studio and verify lms
Download LM Studio only from the official website, open it at least once, then open a new terminal. The lms CLI ships with the application and needs no separate installation.
lms --help lms ls lms ps
Download and load one model
In Discover, choose one chat or instruct variant. Record its size, format, quantization, and license. GGUF variants such as Q4, Q5, and Q8 trade memory against fidelity; Apple Silicon can also use supported MLX models. Do not download several variants just to complete this lesson.
Load the model with conservative settings and ask: Explain local AI in one sentence. Then use lms ps to record the loaded model identifier.
lms ps
Start the local API server
Open Developer and enable Start Server, or use the CLI. The official examples use port 1234. Keep Serve on Local Network disabled for this first test.
lms server start lms server status
Do not bind to 0.0.0.0 for this exercise. LM Studio 0.4.0 and newer can require API tokens, but localhost-only testing does not need network exposure.
Verify /v1/models
curl http://127.0.0.1:1234/v1/models
Copy the exact model id from the JSON response. Do not guess it.
Generate through the API
curl http://127.0.0.1:1234/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "MODEL_IDENTIFIER",
"messages": [{"role": "user", "content": "Explain local AI in one sentence."}],
"temperature": 0.2
}'If /v1/models lists the model and /v1/chat/completions returns content inside choices, you have a local inference layer another application can use.
Why use /v1 endpoints?
LM Studio supports OpenAI-compatible Models, Responses, Chat Completions, Embeddings, and other endpoints. Existing clients can often switch their base URL to http://localhost:1234/v1. Compatibility is endpoint-specific and does not make every remote-provider feature identical.
Local does not automatically mean private
After downloading and loading the model, disconnect the network temporarily and repeat a chat. A response demonstrates offline inference. Model search, downloads, updates, cloud features, web search, or remote MCP servers can still use the network. Evaluate each request's model, URL, tools, and data path.
Fast diagnosis
- Model will not load: choose a smaller model, lighter quantization, or shorter context.
- Connection refused: check lms server status and start the HTTP server.
- Chat works but curl fails: the app can chat while the API server is stopped.
- Port 11434: that is commonly Ollama; these LM Studio examples use 1234.
lms ls lms ps lms server status lms log stream
Save reproducible evidence
Date and operating system: CPU / RAM / GPU: LM Studio version: Model / format / quantization / context: Local chat: PASS / FAIL /v1/models: PASS / FAIL /v1/chat/completions: PASS / FAIL Time to first response: Problem found and next change: Official source checked: