Arena: race local models side by side (and when to use Group Chat)
One model is fine for "rewrite this email." It is a poor substitute for a tiny team.
Why multi-model at all?
Three jobs people actually want:
- Compare — same prompt, side-by-side answers, pick a default without guessing for a week
- Debate — a skeptic answers after the dreamer
- Specialize — coder on the home GPU, travel tutor on-device, editor with a tight style prompt
If you only need one voice forever, stay in a normal thread. Multi-model is for disagreement on purpose.
Honest DIY first
Open two tabs. Paste the same prompt. Compare by eye. Optionally paste Model A's answer into Model B with "critique this."
Works with almost any client. Also: fragmented history, no shared Personas, racing four models for latency is tedious. That is why Arena and Group Chat exist.
Arena — race up to four
Arena is the "stop guessing" button.
- Pick up to four models / providers you already use (on-device, LM Studio, Ollama, Jan, oMLX, optional cloud keys you own).
- Enter one prompt.
- Watch them race — speed and answer quality in one place.
- Keep the winner as your daily default, or note which model wins at which job.
Use Arena when you are choosing a default model, testing a new quant, or checking whether the big home GPU model is actually worth the wait for your prompts.
Group Chat — the room
Group Chat is different energy. Several Personas or models share one conversation. Turn modes, parallel replies, scenarios, even group voice.
Rooms I actually use:
- Brainstorm → critic → checklist
- Fast on-device pass + bigger LM Studio model for "do not ship this"
- Tutor + quizzer + skeptic
Casting is Personas. Rehearsal is Group Chat. Auditions are Arena.
Backends that work
Point participants at whatever you already run:
- On-device GGUF / MLX on the phone
- LM Studio on Wi‑Fi or remote
- Ollama on
:11434 - Jan AI
- oMLX on a networked Apple box
Same privacy story as the rest of LM Mini: local when you are local, encrypted remote when you travel. Docs: Getting started, Group Chat if you want the settings map.
FAQ
Arena vs Group Chat?
Arena = same prompt, compare outputs (and speed). Group Chat = ongoing multi-participant conversation with roles and turn modes.
Do I need Pro?
Core local chat paths stay free. Higher group limits and some premium tools may need Pro — check the in-app account screen.
Can I mix Ollama and LM Studio in one room?
Yes, if each participant points at its server. That is half the point.
The bottom line
Stop clipboard relay races. Use Arena to pick a default. Use Group Chat when you want a cast. Keep the models on hardware you control.
Download LM Mini and race two models you already trust. Winner buys coffee. Loser stays on the phone for airplane mode.