On a laptop: one command, and free chat
Share your laptop with the network and chat on the network for free. Everyone gets 5,000 free tokens a day; with a node you get 20,000 free tokens a day once it passes its first checks (usually under an hour), plus every token your node serves for others. The installer picks a model that suits your memory, installs Ollama, sets up a secure tunnel, and starts the node in the background and at login. It pauses while you're on battery.
First create an API key; the installer asks for it.
macOS and Linux
curl -fsSL https://decentralised.si/install/node.sh | sh
Windows (PowerShell)
irm https://decentralised.si/install/node.ps1 | iex
| Your memory | Model it installs | Download |
|---|---|---|
| under 6 GB | qwen2.5:1.5b | ~1 GB |
| 6–11 GB | llama3.2:3b | ~2 GB |
| 12–23 GB | qwen2.5:7b | ~4.7 GB |
| 24 GB or more | qwen2.5:14b | ~9 GB |
Then open the app and choose Free — community nodes. Free chats are answered on other people's computers, which see the text they answer (never who you are), so keep private details out of them. To choose a different model: curl -fsSL https://decentralised.si/install/node.sh | DSI_MODEL=qwen2.5:7b sh. To remove everything: curl -fsSL https://decentralised.si/install/node.sh | sh -s -- --uninstall (Windows: & ([scriptblock]::Create((irm https://decentralised.si/install/node.ps1))) -Uninstall).
1 · Choose a model for your hardware
What is the most memory your model can use? For a GPU, that's its VRAM. For a Mac or a machine without a GPU, it's the system RAM, minus room for everything else you run.
| Model (Ollama tag) | Download | Good at | Level | Reward |
|---|
Downloads are Ollama's default 4-bit builds (approximate). Each model's level is what dsi-node detects from its name, and it sets how much each unit of work is worth: L0 ×1, L1 ×2, L2 ×4. Serving L2 models needs a 100 PAI provider bond (L3 needs 1,000). Stake is locked, not spent, and can be slashed for dishonesty. New models appear all the time: any model in the Ollama library works; set its tag as DSI_MODEL.
2 · Start your node
qwen2.5:7b (the commands below use it)- Install Docker (Docker Desktop on Mac and Windows).
- Create an API key in the console. Your node's rewards go to that key's organisation.
- Follow the steps for your machine:
First install the NVIDIA Container Toolkit so containers can use the GPU (check with nvidia-smi). Then:
mkdir dsi-node && cd dsi-node
curl -fsSLO https://decentralised.si/dl/docker-compose.yml
curl -fsSLO https://decentralised.si/dl/docker-compose.gpu.yml
export DSI_API_KEY=ds_live_... # your key
export DSI_MODEL=qwen2.5:7b
docker compose -f docker-compose.yml -f docker-compose.gpu.yml up -d
Use Docker Desktop with the WSL 2 backend and an up-to-date NVIDIA driver; GPU support is built in. In PowerShell:
mkdir dsi-node; cd dsi-node
curl.exe -fsSLO https://decentralised.si/dl/docker-compose.yml
curl.exe -fsSLO https://decentralised.si/dl/docker-compose.gpu.yml
$env:DSI_API_KEY = "ds_live_..." # your key
$env:DSI_MODEL = "qwen2.5:7b"
docker compose -f docker-compose.yml -f docker-compose.gpu.yml up -d
Docker on a Mac cannot use the Apple GPU, so run the model with the native Ollama app, which uses Metal and is many times faster. Docker then only runs the tunnel and the node agent.
# 1. model server (native, uses the GPU) brew install ollama # or install the app from ollama.com ollama serve & # not needed if the Ollama app is running ollama pull qwen2.5:7b # 2. tunnel + node agent mkdir dsi-node && cd dsi-node curl -fsSLO https://decentralised.si/dl/docker-compose.yml export DSI_API_KEY=ds_live_... export DSI_MODEL=qwen2.5:7b export LLM_BASE_URL=http://host.docker.internal:11434/v1 docker compose up -d --no-deps cloudflared dsi-node
Keep the Mac awake while serving (System Settings → Energy, or caffeinate -s).
This is for ROCm-supported Radeon and Instinct cards on Linux:
mkdir dsi-node && cd dsi-node curl -fsSLO https://decentralised.si/dl/docker-compose.yml curl -fsSLO https://decentralised.si/dl/docker-compose.rocm.yml export DSI_API_KEY=ds_live_... export DSI_MODEL=qwen2.5:7b docker compose -f docker-compose.yml -f docker-compose.rocm.yml up -d
Any machine with Docker works. Pick a small model (≤ 4 GB): CPUs manage a few tokens per second, which is fine for short answers.
mkdir dsi-node && cd dsi-node curl -fsSLO https://decentralised.si/dl/docker-compose.yml export DSI_API_KEY=ds_live_... export DSI_MODEL=qwen2.5:7b docker compose up -d
What starts
Serves the model on your machine. model-pull downloads DSI_MODEL once.
Gives the node a public https address through a Cloudflare Tunnel. Your IP stays private and no ports are opened.
Registers your model, accepts only requests signed by the router, streams answers, and heartbeats every 30 s.
The compose files and node agent are open source: docker-compose.yml · packages/node · checksums in SHA256SUMS.
3 · Check it's serving
docker compose logs -f dsi-node # [dsi-node] registered as n… with 1 model(s); status: probation until canaries pass # [dsi-node] live at https://….trycloudflare.com/v1 as n… docker compose exec dsi-node node /app/dsi-node.mjs earnings # work units and PAI
Your node now counts on the live network map in your city: nodes online, requests per minute, tokens and projected rewards update every few seconds.
What to expect at first.
- Probation: new nodes start on probation. The router sends test prompts (canaries) to check the node serves the model it claims; after three passes it becomes active.
- Traffic follows reputation: a new node gets little traffic until it has a record of good, fast answers. How much it gets after that depends on demand for its model and level.
- Uptime counts: keep the node online. Nodes that stop heartbeating stop receiving work, and missed service-level targets lower reputation.
Make your node a specialist
A small model can answer like an expert in one field when it has the right notes. Give the node a persona and a folder of Markdown notes: put persona.md and a knowledge/ folder in a notes folder next to docker-compose.yml (it is mounted at /notes), and set these before docker compose up -d. For each question, the node searches the notes and passes the most relevant passages to the model. It also declares its expertise, so routers such as Synapse send it the questions it is best at.
DSI_DOMAINS=law:0.95,law.corporate:0.9 # declared expertise (taxonomy path: competence) DSI_PERSONA_FILE=/notes/persona.md # system prompt for every question DSI_KNOWLEDGE_DIR=/notes/knowledge # Markdown, split at ## headings, searched per question DSI_TRACE_RETRIEVAL=1 # start answers with the passages consulted
Declared expertise is tested like everything else: canaries check the model, and results decide how much traffic the node gets. Two working examples, a law node and an accounting node, are in nodes/expert.
Useful commands
docker compose ps # status docker compose down # stop serving DSI_MODEL=qwen2.5:14b docker compose up -d # switch model (with your -f files) docker compose pull && docker compose up -d # update
Privacy and rewards, plainly
What the map shows
Your node is placed at city level, using Cloudflare's geolocation of the node's IP, rounded to about 10 km. The map shows only per-city totals. Your node's name, address, tunnel URL and account are never public. To stay off the map, set DSI_NODE_LOCATION=hidden; the node then counts only in the "location hidden" total.
What your node sees
Your node sees single requests, signed by the router, with no account, key or user identity, and pseudonymised by the sender's router where they use one. It does see the text it is asked to answer, as any model server does. Don't run a node on a machine you don't trust with that.
How rewards work
Each day the network's emission is split: 70% to providers, pro rata to verified work units. Work units combine tokens, bandwidth, reputation and the model's level. PAI currently lives on an off-chain ledger: no token has been issued, and none is scheduled. Earnings depend on demand and are not guaranteed. See the whitepaper for the exact formulas.