Use it
Room settings
Every setting in the Room settings menu, what it changes and its default.
Open Room settings with the sliders icon at the right of the room’s header. On a phone it rises from the bottom as a sheet. The top of it shows how many devices are in the room.
The Joining, Conversation and Model sections show only on the screen of the device that created the room. Everyone sees the Room section.
Joining
Section titled “Joining”Ask before new devices join
Section titled “Ask before new devices join”On by default. A device that types the room code waits until you press Allow. Your invite link and QR code let a device in without asking. Off, anyone with the code joins. See Invite devices.
Conversation
Section titled “Conversation”Who sees the chat
Section titled “Who sees the chat”Everyone (default), Only me or Whoever asked. It decides which screens show the text, not who computes it. API requests from pooled serve follow it too. See Chat.
Answer length
Section titled “Answer length”| Option | Longest answer |
|---|---|
| Short | About a paragraph at most (150 tokens). |
| Normal (default) | A few paragraphs (400 tokens). |
| Long | Room for long answers and code (1,200 tokens). |
A capped answer can be continued with Continue.
Think before answering
Section titled “Think before answering”Off by default. On, the model reasons before it answers: slower, more careful. The reasoning comes out of a 1,200-token budget shared with the answer.
Changes to answer style and wording apply from the next question. Every screen shows a short notice, for example “answers now: concise”.
Answer style
Section titled “Answer style”A voice for every answer: Plain (default), Concise, Like I’m five, Pirate, Haiku or The room speaks.
Wording
Section titled “Wording”| Option | Sampling | What it does |
|---|---|---|
| Creative (default) | temperature 0.8, top-k 40 | Varied wording: ask twice, get two different answers. |
| Focused | temperature 0.4, top-k 20 | Steadier wording, fewer surprises. |
| Exact | greedy | Always the likeliest word: the same question gets the same answer, on any room shape. |
Layer split
Section titled “Layer split”| Option | What it does |
|---|---|
| For speed (default) | The fastest devices hold the layers, with the fewest hops; the rest join to ask. Before the first answer measures them, computers go before phones. |
| By memory | Every computer holds layers, sized by the memory it gives. A phone holds layers only when the computers can’t fit the model. |
It takes effect when the layers are dealt again, so changing it shows a Re-deal the layers button. Details: How layers are split.
Re-deal on its own
Section titled “Re-deal on its own”Experimental, on by default. When a device drops and is not back within a minute, the host deals its layers to the devices still here. An answer or Code run in progress carries on. Off, the room waits for the device or for you to press Re-deal. See Device drops and recovery.
| Action | What it does |
|---|---|
| Save the chat | Downloads the conversation as a Markdown file. |
| Room card | A picture of this room’s devices and speed, to save or share. |
| Clear cached weights | Frees this device’s disk. Press twice. The next Start downloads again. |
| Cached models list | One row per cached model with its size and a Delete button, so you can remove one model and keep the others. Hidden on phones, which don’t cache weights. |
Below that, a line says whether this device is keeping its screen awake.
Settings in other places
Section titled “Settings in other places”- Memory this device gives: the Memory stepper on the join screen, or This device gives in the room.
- Allow API clients: on the Serve API page, host only. On by default.
- Relay (TURN): the Network box under the join form. See Networks and work Wi-Fi.
The host’s settings are saved with the room, so a host that reloads gets them back.
Options in the page address
Section titled “Options in the page address”A few choices, like context length (?ctx=) and the layer split (?split=), go in the room’s address before you press Start, on the device that hosts the model. See URL options.