SwarmLLM is a free app that gives you a ChatGPT-style assistant running on your own computer, using open AI models from Meta, Google, Alibaba, Mistral and others. When a model is too big for one machine, computers running SwarmLLM team up over the internet and run it together, each doing part of the work.
No account · No subscription · No ads · No crypto · Open source · Updates itself
A giant model has now run on the swarm. Bigger AI models give noticeably smarter answers, but the biggest open ones (70 billion "parameters" and up) need equipment normally found only in data centres — no home computer can hold one. In October 2026, Llama 3.3 70B answered across four ordinary computers in Thailand, Australia and Italy, each holding a part of it. It is slow so far — about a word a second — because those computers are continents apart. The goal now is giant models that are fast, free for everyone, and every computer that joins near another brings that closer.
Download · Join the Discord — find people near you to team up with · For developers
An early version, improving every week. Fixes reach you automatically. Hit a rough edge? Tell us.
Pick the download for your computer from the latest release page.
Windows
- Download
swarmllm-windows-x86_64-gpu.zipif your PC has an NVIDIA graphics card — most gaming PCs do. AMD or Intel graphics, no graphics card, or not sure? Takeswarmllm-windows-x86_64-cpu.zip, which works on any PC. - Right-click the downloaded file and choose Extract All.
- Open the extracted folder and double-click
swarmllm.exe. Windows shows a blue box saying "Windows protected your PC" — it does that for programs from small projects that aren't registered with Microsoft yet. Click More info, then Run anyway. - A black window with scrolling text appears: that is SwarmLLM. Leave it open (minimising is fine) — closing it switches SwarmLLM off. The app opens in your web browser by itself.
Mac (with an Apple chip — M1 or newer)
- Download
swarmllm-macos-aarch64.tar.gzand double-click it to unpack. - Keep the folder in Documents or Downloads — not in Applications, where it can't update itself.
- Double-click
swarmllm. macOS blocks the first launch because the app isn't registered with Apple yet: open System Settings → Privacy & Security, click Open Anyway, and double-click it again. - A Terminal window opens — that's SwarmLLM running, so leave it open. The app appears in your web browser.
Macs with an Intel chip aren't supported yet.
Linux — download swarmllm-linux-x86_64-cuda.tar.gz if you have an NVIDIA
RTX 30-series card or newer, otherwise swarmllm-linux-x86_64.tar.gz, unpack
it and run ./swarmllm run.
Then pick a model and start chatting. The app shows which models fit your
computer. Your first one is a one-time download of a few gigabytes — about the
size of a big game update. If the browser doesn't open by itself, type
localhost:8800 into its address bar.
And then leave it running, and tell someone near you. While SwarmLLM is open, your computer helps the swarm with what it can spare. Every computer nearby makes bigger models faster for everyone in your area — here's why.
Very old computer, Docker or Linux packages?
If SwarmLLM closes straight away with an Illegal instruction message, your
processor is from before 2013 — download the file ending in -baseline
instead. Docker images, .deb/.rpm packages (which run as a background
service and don't update themselves) and step-by-step help for every system
are in the
installation guide.
- Chat with AI on your own computer. Download a model once and it runs on your machine. Unlike a cloud chatbot, there's no company on the other end.
- Use models too big for your computer. Computers in the swarm each hold a part of a bigger model and work through your question together — fastest when they're close to each other, like your own laptop and desktop, or people in the same region. A 70-billion-parameter model has run this way across three countries.
- Help others with what you can spare. While SwarmLLM is open, your computer keeps parts of popular models and helps answer other people's questions — like seeding a torrent. There's no payment and no token; what you get back is the swarm itself: every model it can run, you can use.
- Link your own devices. Put your laptop, desktop and home server into one private group and use them as one bigger machine.
- Plug it into the apps you already use. Chat apps, coding assistants and AI agents that work with ChatGPT or Claude can use SwarmLLM instead — see below.
The app comes in 21 languages, works on a phone, and shows which model gave each answer and how long it took.
An AI model is built from a stack of layers. SwarmLLM cuts a model into parts, and each computer in a team holds a few of them. Your question is passed from one computer to the next like a baton in a relay race — each does its share of the work, and the answer comes back to you. (If you know BitTorrent: it's that idea, applied to running AI instead of sharing files.)
your question ──▶ [ computer A ] ──▶ [ computer B ] ──▶ [ computer C ] ──▶ answer
parts 1–10 parts 11–20 parts 21–30
- No central server. Computers find each other on their own, even behind home routers.
- Nobody has to download the whole model. Each computer keeps only the parts it has room for, within limits you set — or every part, if the model fits and you want to run it yourself.
- Every part is checked. Model parts are verified when they arrive and every time they load, so a bad copy is caught and replaced.
Speed here is tokens a second. A token is a word or part of one, and most people read about 5 a second. Every figure below comes from this project's test computer, a laptop with an 8 GB NVIDIA RTX 3070 graphics card and an 8-core AMD Ryzen 7 5800H processor, alone or teamed up with others, and names the version it was measured on.
On one computer: the v0.3.268 release check, run on the downloaded release. Each model reads a 1,800-token document and writes 128 tokens about it.
| Model | Runs on | Writing (tokens a second) | Reading (tokens a second) |
|---|---|---|---|
| Qwen2.5 Coder 7B | the graphics card | 60–63 | 1,820–1,970 |
| Qwen2.5 14B | card and processor together (too big for the card) | 17.3–18.3 | 690–700 |
| Qwen3 30B-A3B (2-bit) | card and processor together (a mixture of experts, too big for the card) | 20.3–20.5 | 250 |
| Llama 3.2 3B | the processor only | 15.5–15.7 | 70 |
Across computers: the same laptop in Thailand, teamed up with volunteers' computers: one with an RTX 3060 card in Australia, about 400–460 ms there and back, a laptop with an RTX 4050 card in Europe, about 220–240 ms, and for the 70B asked again also a second computer in Thailand and one in Italy. A small model guesses the next few words and the big one checks them all at once.
| Model, and how it is shared | Speed | Version |
|---|---|---|
| Qwen2.5 14B, about half on this laptop and half in Australia | 16–29 writing code, 6.1–10 writing prose | v0.3.261 |
| Qwen2.5 14B, the same, set to "Start and finish on this computer" | 16.6–20 writing code, 6.6–8.7 writing prose | v0.3.268 |
| Qwen2.5 14B, split three ways: this laptop, the laptop in Europe and the computer in Australia | 12–14 writing code, 4.4–4.9 writing prose | v0.3.261 |
| Llama 3.3 70B, this laptop and the computer in Australia, new questions | 1.8–2.6 | v0.3.264–v0.3.265 |
| Llama 3.3 70B, four computers in Thailand, Australia and Italy, a question asked again | 1.9–2.0 | v0.3.255 |
On code, two computers on different continents were faster than this laptop running the same 14B alone (12–14 tokens a second): code is easy to guess, so the guesses keep the long connection busy. Prose is harder to guess, so it gains less. A third computer in the middle made it slower: measured side by side, two computers wrote 16–21 tokens a second of code and 6.1–7.3 of prose, three wrote 12–14 and 4.4–4.9, because every check then crosses three long links and fewer guesses are in flight at once. Each range spans several runs; the across-computer rows were measured on builds of the named version (how).
How big a model the swarm can run is set by who is online. The 70B took the spare room of four computers — and a model that size only answers quickly when those computers are near each other, because your question passes through every one of them for every word of the reply. A team spread across continents works — the 70B showed it — but slowly.
So the swarm needs two things: more computers, and enough of them close together. No amount of code can do that part — only people can.
- Install it, leave it running, and bring someone near you — a friend across town, a housemate, your office. Your friend doesn't need to set anything up: they install it too, and the computers find each other.
- Find people near you on Discord and team up.
- Star this page so others can find it.
Already automatic: computers near each other share out a big model's parts and run it together, and where there aren't enough of them nearby, the swarm teams up computers further away — that is how the 70B ran. Next up: making those teams fast, with the goal of answering faster than one computer alone.
It depends on where the model runs, so here's the straight answer:
- On your own computer — yes. When a model runs entirely on your machine, your conversation doesn't leave it.
- In the public swarm — only partly. Connections between computers are encrypted, so nobody watching the internet can read them. But the computers doing the work for you do see what they're working on, and a determined person running one of them could reconstruct your question. Don't send anything sensitive through the public swarm.
- Start and finish on this computer — harder to read, still not private. When your computer holds a model's first and last parts, it does the first and last steps itself (automatically; switch it per model on the model's card on the Dashboard). Other computers then never receive your typed text or the reply — only the numbers in between. Those numbers can still reveal much of what you wrote, so it is a real improvement but not a guarantee, and replies get slower.
- Private Mode keeps your questions on your own linked devices and, by
default, on other SwarmLLM computers it treats as local: any on your network,
or any that answer in under about 2.5 milliseconds. In a shared building or on a fast
city network that can include strangers —
private_mode_allow_lan = falsein the config file turns that part off. Your computer still helps others with what it can spare. Switch Private Mode on under More → My Devices in the app.
The model list in the app shows which models are on this computer. The details are in the security guide.
What's the catch? There's nothing to buy — no account, no ads, no cryptocurrency. The only cost is your own electricity and internet while your computer helps others. It's a community project: the app is free and open source, and the computing power comes from people who leave it running.
How is this different from Ollama or LM Studio? They run AI on one computer, and they're great at it. SwarmLLM does that too — and also lets computers pool together, your own devices or the public swarm, to run models one machine can't hold. (How it compares with Petals, exo and others.)
What do I need? A Windows, Linux or Apple-chip Mac computer, ideally with 8 GB of memory or more, and a few gigabytes of free disk space per model. A graphics card makes it much faster but isn't required: it needs an NVIDIA RTX 30-series or newer, on Windows and Linux alike, and any other card leaves the work to the processor. On a Mac it runs on the main processor for now — slower than a PC with a gaming card, but a MacBook Air can run the smaller models.
Will it slow my computer down? It can, while it's answering someone. SwarmLLM keeps a slice of your graphics card free and fits around whatever memory other programs are already using, but it doesn't know when you're gaming or busy. If something feels slower, close the SwarmLLM window while you play, pick a lighter contribution level in Settings, or set a Resource Schedule so it holds back at the times you choose. It only runs when you open it; it doesn't start with your computer.
How much internet does it use? A computer left running stays in touch with the swarm all the time, and uses more while it shares model parts or answers requests. Measured on 24 September 2026 on a computer running the current release, idle and connected to six others: about 110 MB an hour received and 130 MB an hour sent — roughly 5 to 6 GB a day if it runs all day. Most of that is the swarm telling each other which models exist. That is still too much for a capped or pay-per-gigabyte connection, and we are still bringing it down. You can cap how fast it shares model parts in Settings, and it uses nothing while it is closed.
How much disk space? SwarmLLM keeps what it stores within a disk limit — 50 GB by default — which you can change in Settings. Whatever the limit, it always leaves at least a tenth of your disk free.
Is it a virus? No. Every line of the program is public for anyone to inspect, and SwarmLLM's automatic updates only install releases signed by the project. Windows and macOS warn you the first time only because the app isn't registered with Microsoft or Apple yet — common for small free projects. Other computers in the swarm can never run programs on yours: they only exchange parts of AI models and the numbers the AI works with.
How do I remove it? Close the SwarmLLM window, then delete the folder you
unpacked and the folder where it keeps models: %APPDATA%\swarmllm on Windows,
~/Library/Application Support/swarmllm on a Mac, ~/.local/share/swarmllm
on Linux.
Which models can I run? Popular open models — Llama, Qwen, Mistral, Gemma, Phi, GLM and more, from small ones that run on a laptop up to a 70-billion-parameter Llama shared across the swarm. The app shows which ones your computer, or the swarm, can actually run.
(You can skip this part if you just want to chat.)
A free OpenAI- and Anthropic-compatible endpoint on your own machine. Point
any agent, coding assistant or chat UI at http://localhost:8800 — no
per-token bill, and models bigger than your machine behind the same URL. Copy
your key from Settings → Access Token in the app:
export SWARMLLM_KEY=... # from Settings → Access Token
curl http://localhost:8800/v1/chat/completions \
-H "Authorization: Bearer $SWARMLLM_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "llama-3.2-3b-instruct-q4-k-m",
"messages": [{"role": "user", "content": "Hello!"}]}'As a Claude Code backend:
ANTHROPIC_BASE_URL=http://localhost:8800 ANTHROPIC_AUTH_TOKEN="$SWARMLLM_KEY" \
claude --model qwen2.5-coder-7b-instruct-q4-k-mAlso included: tool calling with local models, streaming, an MCP server, and
optional routing to 12 cloud providers with your own keys. Agents send long
prompts, so raise the context size first (max_seq_len_override = 32768
under [inference] in config.toml). That setting covers the parts of a model
your computer runs; a part running on another computer follows that computer's
own setting, and the swarm routes around one that is set too short.
- API reference — OpenAI, Anthropic, Responses and MCP
- OpenClaw setup — a plugin that adds SwarmLLM to OpenClaw's setup wizard
- Configuration and command line
- Architecture (every document, one page), performance and how it compares
To build it yourself:
# Requires Rust 1.90+
git clone https://github.com/enapt/SwarmLLM.git && cd SwarmLLM
cargo build --releaseGPU builds — the build command for each release variant — are in CONTRIBUTING.md.
- Discord — find people to team up with, ask questions, get help
- Issues — bugs and feature requests (
swarmllm diagnosticsprints a report that's safe to paste; what else helps) - Discussions — ideas and questions
- Changelog — what changed recently
- Security — please report vulnerabilities privately
- Contributing — build, test and send a pull request
SwarmLLM is built by a human developer working with Claude Code, an AI programming assistant: the human sets the direction, tests and reviews, and Claude writes the code. We say so openly so you can judge the project on its merits. Every change is checked by thousands of automated tests (4273 lib tests + 80 integration tests) before it ships, and regular automated audits are logged in the repository. Scrutiny and contributions are welcome.
Dual-licensed under MIT and Apache 2.0.