Appearance
CLI reference
Every Diffuse command, its flags, and defaults.
bash
diffuse <command> [options]
diffuse --version
diffuse --helpchat
Interactive chat with a model on the network.
| Flag | Default | Meaning |
|---|---|---|
--bootstrap <urls> | built-in sentinels | sentinels to discover the network |
--memory | off | keep conversation history across turns |
--max-tokens <n> | 512 | longest answer per turn |
Inside a session, /attach, /image, /audio and /video send a file with the next message, and a model that answers with audio or pixels writes its answer to a file. The full command list is in the chat guide.
query
Ask one question, print the answer, exit.
| Flag | Default | Meaning |
|---|---|---|
--model <id> | required | model to query |
--prompt <text> | required | prompt to send |
--image <path> | none | image to send with the prompt, repeatable |
--audio <path> | none | audio clip to send, repeatable |
--video <path> | none | video to send, repeatable |
--media <path> | none | any attachment; kind read from the extension |
--bootstrap <urls> | built-in sentinels | sentinels to discover the network |
--max-tokens <n> | 80 | maximum tokens to generate |
--steps <n> | 20 | denoising steps, for a model that answers by diffusion |
--patches <n> | 1 | pieces each denoising step is cut into across the nodes |
--seed <n> | 0 | same seed and prompt give the same answer |
Attachments are consumed on your machine; only activations leave. When the model answers with something other than text, the file is written to the current directory and its path printed. See images, audio and video.
--steps, --patches and --seed apply only to models that answer by diffusion; see diffusion across nodes.
models
List the models currently hosted on the network.
| Flag | Default | Meaning |
|---|---|---|
--bootstrap <urls> | built-in sentinels | sentinels to discover the network |
serve
Run a local OpenAI-compatible HTTP server in front of the network.
| Flag | Default | Meaning |
|---|---|---|
--port <n> | 8080 | port to listen on |
--host <addr> | 127.0.0.1 | bind address, loopback only by default |
--model <id> | none | default model when a request omits one |
--bootstrap <urls> | built-in sentinels | sentinels to discover the network |
host
Join the network: profile, pick a slice, load it, announce, and serve.
| Flag | Default | Meaning |
|---|---|---|
--model <id> | optional | model to serve a slice of; omit to open the marketplace |
--worker <url> | http://127.0.0.1:50051 | worker endpoint |
--listen <addr> | 0.0.0.0:9440 | gossip listen address |
--bootstrap <urls> | built-in sentinels | sentinels to join through |
--overhead <f> | 0.3 | memory fraction held back as headroom |
--public-addr <addr> | auto | address to advertise to peers |
plan
Analyze this machine and show which slice it would host, without joining.
| Flag | Default | Meaning |
|---|---|---|
--model <id> | required | model to analyze against |
--worker <url> | http://127.0.0.1:50051 | worker endpoint |
--overhead <f> | 0.3 | memory fraction held back as headroom |
--bootstrap <urls> | none | sentinels for a network view |
Environment
| Variable | Default | Meaning |
|---|---|---|
HF_TOKEN | none | Hugging Face token; unlocks gated models, stays on your machine |
DIFFUSE_WORKER_DIR | installed copy, then ./worker | which worker to run; see troubleshooting |
DIFFUSE_WORKER_PORT | 50051 | port the worker listens on |
DIFFUSE_WORKER_DEVICE | auto | cpu, cuda, or mps; detected when unset |
DIFFUSE_WORKER_MAX_BATCH | 8 | concurrent requests grouped into one pass; 1 disables |
DIFFUSE_WORKER_CONCURRENCY | 16 | how many requests the worker serves at once |
DIFFUSE_WORKER_THREADS | torch default | CPU threads for inference |
DIFFUSE_WORKER_CACHE | Hugging Face default | where model weights are stored |
DIFFUSE_ASCII | unset | plain symbols instead of Unicode, for limited terminals |
RUST_LOG | warn | info shows routing and timing, debug shows every hop |