Group chat
Multi-model AI rooms with Personas, turn modes, parallel streaming, scenarios, and group voice in LM Mini.
Answer first
Group chat puts multiple AI participants in one thread. Each can use the same model with different roles — or different models and providers — with its own name, prompt, tools, and voice.
Overview: Group Chat: Multiple AI Models in One Conversation.
Availability
Group Chat is in beta. New users get a free preview (up to 2 participants). Pro raises the limit (up to 5) and keeps the feature after the preview window.
Why use it
- Same model, many personalities (writer’s room, debate)
- Different models in one transcript (compare tone and reasoning)
- Reuse Personas (including built-in Mini)
- Per-participant tools — researchers search; narrators stay text-only
- Group voice speaks replies in order with each Kokoro voice
Start a group
- From home, ⋯ → New Group Chat (Mac: rail / overflow where available)
- Add Personas or Add Model participants (one model can play many roles)
- Configure each: name, binding, avatar, color, prompt, tools, voice
- Room rules: turn mode, reply-to-user-only, parallel streaming, auto load/unload
- Optional: your display name + scenario so everyone shares the premise
- Start — bindings are checked first (no silent fallback to a random backend)
Participant options
| Setting | What it does |
|---|---|
| Display name | Label in transcript and Ask bar |
| Model binding | Provider + model for that participant |
| System prompt | Role and style for that participant only |
| Avatar / accent | Visual separation |
| Tool calling | MCP / built-in tools on or off |
| Voice | Kokoro speaker in group voice |
Turn modes
- Take turns (round robin) — ordered replies after you send
- Talk when mentioned (autonomous) — join when @mentioned / room rules say so
- I choose (manual) — Ask: chips; tap who speaks next
- Reply to user only — less AI-to-AI chatter
- Parallel streaming — several replies at once
- Auto load / unload — “Save memory on local models” for VRAM
Scenarios
Scenario text stays with the room so interview / writers’ room / debate premises do not drift.
Daily workflows
- Writer’s room — ideate / critique / edit
- Product review — PM, engineer, marketer, skeptic
- Interview practice — interviewer + coach + rewriter
- Model comparison — same prompt, different engines, one thread