Arena
Race up to four models on the same prompt to compare speed and answer quality in LM Mini Arena.
Arena runs the same prompt across multiple models at once so you can compare tokens/sec, time-to-first-token, and the written answers — then pick a default for chat or Personas.
Open Arena
- Phone: home ⋯ → Arena
- Mac Home: Arena in the nav rail (⌘4)
- Welcome wizard: skippable Find a model that fits (about a minute)
What you can race
Add contestants from this device, LM Studio, Ollama, JAN AI, oMLX, Unsloth, or cloud (cloud contestants are Pro). Free races use the built-in prompts with up to four models. Pro unlocks your own prompts, cloud contestants, and saved race history.
Suggested lanes: this device vs desktop, “Race top 2”, Quick vs Full suite — pick what you have loaded.
How a race works
- Pick contestants and a prompt (or use the built-in discovery prompts)
- Start — streams run live with speed metrics
- Read answers, note TTFT / tok/s
- Optionally apply that model as your everyday default or Persona binding
Anonymous benchmarks
Free races share anonymous speed results — model, device class, and speed numbers, never your prompts or answers. You confirm this once before your first race. Pro users can turn sharing off under Settings → Advanced → Share Arena speed results.