---

name: lm-studio description: This skill should be used when the user asks to "run local LLMs", "use LM Studio", "configure local AI server", "estimate VRAM requirements", "load a model locally", or needs guidance on OpenAI-compatible local API usage, model quantization selection, GPU offload configuration, MCP server integration, or headless LLM server management. Covers local AI inference, CLI automation, SDK integration, and hardware optimization. version: 1.0.0 license: MIT

LM Studio Local LLM Integration

Comprehensive integration for running AI models locally via LM Studio.

Quick Reference

Server Endpoints:

• Base URL: http://localhost:1234/v1
• Chat Completions: POST /v1/chat/completions
• Embeddings: POST /v1/embeddings
• List Models: GET /v1/models
• Responses (Stateful): POST /v1/responses

Default Port: 1234 (configurable via --port)

---

API Endpoints

LM Studio provides an OpenAI-compatible HTTP server.

Authentication: API key not required but some clients expect non-empty string (use "lm-studio").

---

CLI Commands Reference

The lms CLI enables headless operation and automation.

Server Management

Copy# Start server with CORS enabled
lms server start --port 1234 --cors

# Stop server
lms server stop

# Check status
lms server status

Model Management

Copy# List downloaded models
lms ls
lms ls --json  # JSON output for scripts

# List loaded models
lms ps
lms ps --json

# Download model
lms get deepseek-r1
lms get <huggingface-url>
lms get --mlx <model>  # Apple MLX optimized

# Load model with configuration
lms load <model> --ttl 300 --gpu auto --context-length 8192

# Estimate VRAM before loading
lms load --estimate-only <model>

Logging & Monitoring

Copy# Stream server logs
lms log stream --source server

# Stream inference logs with stats
lms log stream --source model --filter input,output --stats --json

---

Model Configuration Parameters

Inference Parameters

Load-Time Configuration

---

VRAM Requirements Guide

Quantization Levels

• Q4_K_M: Best balance for consumer hardware (recommended)
• Q8_0: Higher quality, ~2x VRAM of Q4
• FP16/BF16: Full precision, research/workstation only

Requirements by Model Size

Rule: If model fits entirely in VRAM, runs fastest. Offloading to RAM is up to 30x slower.

Use scripts/check_vram.py to estimate requirements before loading.

---

Recommended Models by Use Case

Reasoning & Logic

DeepSeek-R1 Distillations:

• deepseek-r1-distill-qwen-7b, deepseek-r1-distill-qwen-32b
• Temperature: 0.5-0.7
• No system prompt (instructions in user message)
• Math directive: "Please reason step by step, and put your final answer within \boxed{}."

Nemotron 3 Nano:

• nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-GGUF
• Temperature: 1.0, Top P: 1.0
• Use /think or /no_think control tokens

Coding & Agentic Workflows

Qwen3-Coder:

• qwen3-coder-30b-a3b-instruct
• Temperature: 0.3
• Context: up to 256k tokens

GLM-4.7:

• zai-org/GLM-4.7
• Excellent for web page generation and multi-step tool use

Vision (Multimodal)

Qwen2.5-VL/Qwen3-VL:

• qwen2.5-vl-7b-instruct, qwen3-vl-8b-instruct
• Set image resize bounds to 2048px minimum

LFM2.5-VL:

• LiquidAI/LFM2.5-VL-1.6B
• Optimized for edge devices and laptops

---

MCP Server Configuration

LM Studio acts as an MCP Host, connecting to external tool servers.

Configuration File: ~/.lmstudio/mcp.json

Remote Server Example

Copy{
  "mcpServers": {
    "hf-mcp-server": {
      "url": "https://huggingface.co/mcp",
      "headers": {
        "Authorization": "Bearer <YOUR_HF_TOKEN>"
      }
    }
  }
}

Local Server Example

Copy{
  "mcpServers": {
    "sandbox": {
      "command": "sandbox-server-stdio",
      "args": []
    },
    "my-local-server": {
      "url": "http://localhost:8000"
    }
  }
}

Tool Choice Options:

• "auto": Model decides whether to call tools
• "required": Force tool usage
• "none": Disable tools

---

Python/OpenAI Client Integration

Copyfrom openai import OpenAI

client = OpenAI(
    base_url="http://localhost:1234/v1",
    api_key="lm-studio"  # Placeholder required by some clients
)

completion = client.chat.completions.create(
    model="qwen2.5-7b-instruct",
    messages=[{"role": "user", "content": "Hello!"}],
    temperature=0.7,
    max_tokens=2048
)

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

With Streaming

Copystream = client.chat.completions.create(
    model="deepseek-r1-distill-qwen-7b",
    messages=[{"role": "user", "content": "Explain quantum computing"}],
    stream=True
)

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

---

TypeScript SDK Usage

Copyimport { LMStudioClient } from "@lmstudio/sdk";

const client = new LMStudioClient();

// Load model with configuration
const model = await client.llm.load("qwen2.5-7b-instruct", {
  config: {
    contextLength: 8192,
    gpu: { ratio: 1.0 },
    flashAttention: true,
    useFp16ForKVCache: true
  }
});

// Run inference
const response = await model.respond([
  { role: "user", content: "Write a Python function to sort a list" }
]);

---

Framework Integration

LangChain

Copyfrom langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    base_url="http://localhost:1234/v1",
    api_key="lm-studio",
    model="qwen2.5-7b-instruct"
)

AutoGen

Copyconfig_list = [
    {
        "model": "qwen2.5-7b-instruct",
        "base_url": "http://localhost:1234/v1",
        "api_key": "lm-studio"
    }
]

assistant = autogen.AssistantAgent(
    name="Assistant",
    llm_config={"config_list": config_list}
)

---

System Requirements

macOS (Apple Silicon):

• M1/M2/M3/M4 chip required (Intel not supported)
• macOS 13.4+ (14.0+ for MLX backend)
• 16GB+ unified memory recommended

Windows:

• x64 with AVX2 or ARM64 (Snapdragon X Elite)
• 16GB+ RAM, 4GB+ dedicated VRAM minimum
• NVIDIA RTX preferred (CUDA 12.8), AMD via ROCm

Linux:

• Ubuntu 20.04+ (AppImage)
• x64 with AVX2, 16GB+ RAM, 8GB+ VRAM

---

Performance Optimization

1. Enable Flash Attention for NVIDIA RTX GPUs
2. Use Q4_K_M quantization for best speed/quality balance
3. Set appropriate context length (smaller = faster, less VRAM)
4. Enable FP16 KV Cache to reduce memory usage
5. Configure Idle TTL for automatic resource cleanup
6. Use speculative decoding with draft models for 2-3x speedup

Speculative Decoding Example

Copy{
  "model": "deepseek-r1-distill-qwen-7b",
  "draft_model": "deepseek-r1-distill-qwen-0.5b",
  "messages": [...]
}

---

Bundled Resources

• scripts/check_vram.py - Estimate VRAM requirements
• scripts/server_health.py - Monitor server status
• scripts/model_benchmark.py - Performance testing
• references/integration-patterns.md - RAG, multi-model, IDE integration examples