Chat with the network

diffuse chat is the interactive way to talk to a model on the network.

diffuse chat

It connects through the sentinels, picks a servable model (prompting you if there is more than one), starts a local tokenizer, and streams the reply token by token. Type a message and press enter. Type /quit to leave.

Options

Flag	Default	Meaning
--bootstrap <urls>	built-in sentinels	comma-separated sentinels to discover the network
--memory	off	keep conversation history across turns
--max-tokens <n>	512	longest answer per turn

--max-tokens means two different things depending on the model, and the difference matters. A model that answers with words stops when it has finished, so the number is a ceiling it usually never reaches. A model that answers with audio or pixels has no equivalent of "finished": the number is the length of the answer, and every step is a round trip through the network. On MusicGen, 512 steps is about ten seconds of music and several minutes of waiting. Start small.

Commands

Command	What it does
/help	list these
/attach <file>	attach a file; its kind comes from the extension
/image <file> · /audio <file> · /video <file>	attach, saying what it is
/files	show what is attached to the next message
/detach	drop the attachments
/reset	clear conversation memory
/clear	clear the screen
/stats	turns and tokens so far
/save [file]	write the transcript as markdown
/quit	leave

Sending pictures, sound and video

Attach a file, then ask your question. The attachment applies to the next message only, and is dropped once the answer arrives.

● › /image ~/photos/receipt.jpg
  ◆ attached image   receipt.jpg · 240 KB
● › What is the total on this receipt?

The file never leaves your machine. It is read here, turned into hidden states by the part of the model that lives on your side, and only those activations travel — the same thing that would travel for a text prompt. See multimodal for which models accept what.

If you attach something the model cannot read, the turn fails with a message saying so rather than quietly answering as if no file had been sent.

Answers that are not words

Some models answer with audio or an image rather than text. Chat handles those the same way you would expect: it says what the model produces, generates it, writes the file next to you, and prints the path.

● › a calm piano melody
  → this model answers with audio on 4 streams
  → generating over encrypted channel...

  answer:
  36.3 KB of audio/wav after 31 steps
  /home/you/diffuse-0000019f.wav

The transcript records the path, so /save keeps a usable record of what was produced.

Memory

By default each message is standalone: the model does not see earlier turns. Pass --memory to carry the full history into each turn, so the conversation builds on itself.

diffuse chat --memory

Without memory, a session is a series of independent questions. With it, it is a continuous conversation. History lives only in your local process and is never stored.

Joining a specific network

To talk to a private network or your own sentinel instead of the public one:

diffuse chat --bootstrap http://your-sentinel:9440

What stays local

Your machine runs the tokenizer, drives the generation loop, and decodes the result. Traffic to the network is encrypted end to end. For exactly what a serving node can and cannot see, read the privacy and threat model.

Next

• Ask one question
• Choosing a model