Group Chat: Multiple AI Models in One Conversation
In the Personas post I promised Group Chat next. Here it is. I am a man of my word. Occasionally.
Personas are casting. Group Chat is rehearsal. Same stage, more actors, louder opinions.
What Group Chat actually is
Most AI apps give you one model, one reply, one vibe. Fine for "what's 12 times 8." Less fine when you want a second opinion without opening three apps and copy-pasting. Clipboard promotion incoming.
Group Chat puts multiple AI participants in one conversation. Each can be a Persona you already built, or a model from on-device, LM Studio, or Ollama. You talk. They talk. Sometimes they talk to each other. Sometimes they all answer you at once. You pick the mode.
It feels less like "chatbot" and more like a group text where everyone is weirdly helpful and nobody asks to split the bill.
Why I built it
I kept doing this dance myself: ask Model A for a plan, paste it into Model B for critique, then ask Model C to make it shorter. Useful. Also exhausting. My clipboard deserved a raise.
So Group Chat became the feature request I could not ignore. People wanted a room, not a relay race. And once Personas existed, the cast was already there. Putting them in the same thread was the obvious next step. Community asks. I ship. We iterate. Same loop as voice and Kokoro.
How it feels in practice
A few ways I actually use it:
- Brainstorm room. One Persona generates wild ideas. Another kills the bad ones. A third turns survivors into a checklist.
- Code review panel. Fast on-device model for a first pass. Bigger LM Studio model for the serious "please do not ship this" feedback.
- Study group. Tutor explains. Quizzer tests you. Skeptic asks "but why?" until you really get it.
- Writer's table. One voice drafts. One edits tone. One fact-checks. You stay the editor-in-chief with thumbs that hurt.
You can also run scenarios: a setup where the group already knows the scene. Less "hello everyone," more "you are already in the middle of the problem."
Turn modes (without the boring manual energy)
I will keep this short. The docs go deeper.
- One at a time. Classic round robin. Cleaner to read. Good when you want a conversation, not a choir.
- Parallel. Several models answer the same prompt together. Perfect for comparing quality and speed. Slightly chaotic. Worth it.
- Directed. You pick who speaks next. Useful when you want the critic after the dreamer, not during.
There is also group voice if you like talking out loud. It is the same voice stack you already know, just with more characters trying to get a word in. Headphones recommended. Neighbors optional.
How to try it (rough path)
- Build two or three Personas first if you have not already. Group Chat is more fun with a cast than with three blank "Assistant" clones.
- Open Group Chat / create a multi-model room (wording can vary slightly by app version).
- Add your Personas or models. Mix on-device and desktop if you want. Yes, that means your phone and your home GPU can argue in the same thread. Democracy.
- Pick a turn mode. Start simple: one at a time.
- Ask something that benefits from disagreement. "Give me three product names and roast the weak ones" works better than "hi."
- If the room feels noisy, remove a participant. Group Chat is a dial, not a dare.
Group Chat vs Arena (quick map)
Arena is a race: same prompt, up to four models, compare who is fast and who is smart. Group Chat is a conversation over time. Use Arena when you are shopping for a model. Use Group Chat when you want collaborators.
Privacy, still the point
Nothing about Group Chat sends your writers' room to a random cloud just because there are more speakers. On-device stays on-device. LM Studio and Ollama stay on your network or through Share with phone on Mac (encrypted key only you control) when you are remote. More Personas does not mean more companies reading your drafts.
The bottom line
One model is a tool. A few models in one thread is a tiny team. Group Chat is how LM Mini lets that team live on your phone without turning your life into someone else's training data.
Go make a room. Put your calm editor next to your chaotic idea generator. Let them fight. You take the good parts.
Next up: why private AI still matters in 2026 (yes, even if your only secret is a grocery list and a half-finished novel). Until then, assemble the cast.