Appearance
FAQ
Straight answers to the questions people actually ask.
General
Is Diffuse free?
Yes. It is open source under AGPL-3.0. There is no account, no API key, and no paid tier. You run a binary and you are on the network.
Do I need a GPU?
No. The worker runs on CPU by default and works on ordinary machines, including laptops and a Raspberry Pi. A GPU makes a node faster but is not required.
Is it fast?
No, and it does not pretend to be. A model split across ordinary machines over the public internet is slower than a single datacenter GPU. On the measured three-node run, generation ran at roughly 1.2 tokens per second. The point of Diffuse is privacy and decentralization, not speed. See the benchmarks.
Which models work?
Ordinary chat models, models that read images, audio or video, and encoder-decoders such as Whisper. A few also answer with sound rather than words. What decides is the shape of the computation, not a list of approved names: see what Diffuse can run.
U-Net diffusion, recurrent stacks such as Mamba, and checkpoints that ship their own Python do not work. Mixture-of-experts is unproven.
Privacy
Can a node read my prompt?
On the default path, the node that runs the first slice receives your token ids, which are your prompt. Traffic is encrypted in transit, so the network cannot read it, but the entry node can. A private mode that keeps the first layers on your machine is in progress. Read the honest threat model before trusting it with anything sensitive.
Does Diffuse hide that I am using it?
No. It hides what you say on the wire, not that you are talking. Your IP is visible to the node you connect to, exactly as with any network service.
Are my prompts logged anywhere?
There is no central server, so nothing is logged centrally. A malicious entry node could log what it receives, which is why strong privacy needs local first layers.
Running a node
What does hosting cost me?
Bandwidth, some memory, and CPU while a request flows through your slice. You choose how much by the size of the slice you take, which the planner fits to your machine.
Can I run a node behind my home router?
Yes. If your node cannot accept inbound connections, it serves through an encrypted relay automatically.
Will hosting download the whole model?
No, only the shards that contain your slice, though shard boundaries mean you may download one or two shards. See choosing a model.
How do I keep a node running after I close the terminal?
bash
nohup diffuse host --model Qwen/Qwen2.5-0.5B-Instruct > ~/.diffuse/host.log 2>&1 &Using it
Can I use it with my existing chat app?
Yes. Run diffuse serve and point any OpenAI-compatible client (LibreChat, Open WebUI, Continue, the SDKs) at http://localhost:8080/v1. See the server guide.
Can I run my own private network?
Yes. Point your nodes and clients at your own sentinel with --bootstrap. See self-hosting a network.
Why did my query fail with "model is incomplete"?
Some layer range of that model is held by nobody, so it cannot be served end to end. The error names the missing slices. Wait for them to come online or host one yourself.
Troubleshooting
The command hangs or the worker will not start.
The Python worker imports torch before opening its port. Diffuse now retries for up to 30 seconds and prints a clear error naming the cause. See troubleshooting.
A second run fails to bind a port.
Only one node or client can use the fixed ports per machine. Stop the previous process. See troubleshooting.