llama-swap: Hot-Swap Between Local LLMs Like It's Nothing

llama-swap is a game-changer for anyone running local LLMs. It lets you spin up multiple models and switch between them instantly without restarting anything.

Why This Matters

Testing different models is painful. Stopping one server, starting another, changing your config, repeating. llama-swap eliminates all of that. One endpoint, multiple models.

What You Get

Single binary. One config. Multiple local models.

• llama.cpp / llama-server
• vLLM
• TabbyAPI
• Any OpenAI-compatible server
• Web UI included
• Auto-unload after timeout
• Concurrent models
• Docker support

Installation

# macOS ARM64
curl -L -o llama-swap.tar.gz "https://github.com/mostlygeek/llama-swap/releases/download/v203/llama-swap_203_darwin_arm64.tar.gz"
tar -xzf llama-swap.tar.gz
chmod +x llama-swap

# Linux x86_64
curl -L -o llama-swap.tar.gz "https://github.com/mostlygeek/llama-swap/releases/download/v203/llama-swap_203_linux_x86_64.tar.gz"
tar -xzf llama-swap.tar.gz

Configuration

Create config.yaml:

upstreams:
  ollama:
    url: http://localhost:11434
    type: openai
  
  lmstudio:
    url: http://localhost:1234/v1
    type: openai
    
  vllm:
    url: http://localhost:8000/v1
    type: openai

default: ollama

# Optional: auto-unload idle models
idle_timeout: 300

# Optional: API key protection
api_keys:
  ollama: "your-key-here"

Running It

./llama-swap serve

Web UI: http://localhost:8080/ui
API: http://localhost:8080/v1/chat/completions

Switching Models on the Fly

# Via API
curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "lmstudio:codellama",
    "messages": [{"role": "user", "content": "hello"}]
  }'

# Or use the web UI dropdown

Use Cases

1. Quick tests - OLLAMA for fast iterations
2. Code review - LM Studio with CodeLlama
3. Production - vLLM with big models
4. Comparison - See same prompt, different models

Integration with OpenClaw

In your OpenClaw config, point to llama-swap:

{
  "model": "ollama",
  "base_url": "http://localhost:8080/v1"
}

Swap models via environment or API without touching OpenClaw config.

Compared to Alternatives

Troubleshooting

Port conflicts: Change in config:

server:
  port: 8081

Model not loading: Check URL is correct and server is running

API errors: Verify model names match what’s in your config

Updated 2026-04-23 - More fleshed out guide as requested