Your AI, on your devices.
Brahma spreads open-weight AI models across the Macs, PCs, Raspberry Pis and phones people already own. The node is free, forever. Every node that joins makes the mesh stronger.
How it works
Like Napster, for AI compute.
A Mixture-of-Experts model only needs a few of its experts for each token. Brahma splits the model into small pieces, spreads them over ordinary machines, and brings the experts your prompt needs to you. No GPU, no datacenter.
Install a node
One command on a Mac, PC or Linux box. It sizes itself to the machine.
./install.sh
Join your machines
Add a laptop, a mini PC or a Raspberry Pi with one key. They share the model's experts.
brahma join <key>
Ask, privately
On your own devices, your prompt and its reply never leave your machines.
brahma open
Next: a public mesh, where your node can help run models for others.
Privacy, security and scale
From one laptop to the mothership.
One Brahma node on your laptop.
Install it and your laptop is a node. It holds the model’s experts and can answer on its own.
Your words stay where you type them
The chat runs on the machine in front of you. Other nodes receive only numbers, the vectors an expert needs to do its part, never your text. A private request goes only to machines you mark as trusted.
Vectors aren’t plain text, but they can leak a little, so private mode keeps them on your own trusted machines.
Only machines you approve can join
Machines in one home find each other with signed announcements and talk over encrypted links tied to your mesh key. A machine elsewhere, like your parents' PC, joins only after you add its key, and the link is pinned to that key. A node with an unknown key is turned away.
Commands: brahma join <key> at home, brahma peer add for a machine elsewhere.
Every machine adds capacity
Each node holds some of the model's experts. Add a machine and the mesh can hold more of the model and serve more chats: 24 cloud CPUs served 260 tokens/s together, and a 235-billion-parameter model ran across 9 of them.
The hosted mothership at mybrahma.com is planned, not live.
What's in the beta
Everything here works today.
Private and secure
- Your words stay put
- Prompt and reply text stay on the node running the chat. Only number vectors travel.
- Private mode by default
- Only your trusted machines do the work, and every reply shows which machines computed it and their share.
- Locked links
- Mutual TLS on your LAN, keyed by your mesh key. Internet peers need a key you allowlisted, over TLS pinned to it, revocable at once.
- Safe model files
- Weights are plain data that can't run code. An expert sent between machines is checked against the model's signed release.
Senses and context
- Voice, on device
- Speech to text, text to speech and voice detection run on your machine, off until you turn them on.
- Eyes
- Local vision models turn the camera into a text scene report. People are counted, never identified.
- Attachments
- PDFs, images and audio become text on your node. The files never leave it.
- Connectors
- Read-only context from a curated catalog of MCP servers.
A mesh that runs itself
- No central server
- 37 machines formed one mesh in 4 seconds on AWS.
- Any recent CPU
- Apple Silicon, AWS Graviton, Intel Xeon and AMD EPYC, mixed in one mesh.
- Distance-aware
- Machines in six regions on four continents, up to 217 ms apart: one chat kept 98% of its speed.
- Self-healing
- When a machine leaves, the others pick up its experts.
AI you can inspect
- Every reply explains itself
- Tokens, speed, time to first token, machines and expert calls.
- The explorer
- Which heads, experts and neurons fired, token by token.
- Live view and Network page
- Generation step by step, and a globe of every machine with its load and latency.
- A model that answers questions about itself
- Query the model as a graph; the reasoner shows each step.
A first look
Recorded on real machines. Not a mock-up.
Real Brahma nodes, running on CPUs alone: a MacBook with Qwen1.5-MoE-A2.7B, and a 14-node test mesh for the network view. This is what beta testers install on their own machines.
Chat that stays on your machine
Ask anything and the answer is worked out on your own computer. Under every reply you can see how it was made: how fast it was, how many machines helped, and that it stayed private.
See why every word was chosen
Type “The capital of France is” and watch it land on “Paris”. Every dot is a piece of the model, and the ones that light up are the ones that did the thinking for that word.
Watch the mesh work, live
Slow a reply right down and follow each step: your prompt is read, the right experts are picked, and the next word comes back. The heatmap shows which experts the model leans on most.
Every machine in your mesh, on a globe
Send a prompt and the links light up as experts on other machines pitch in, with the calls per second for each. New machines appear the moment they join.
Ask the model about itself
Brahma keeps a map of the model it runs, so you can ask about any part of it: which layer it sits in, what it reads, and which machine it runs on.
A reasoner that shows its working
Ask a yes-or-no question about the model and the reasoner works it out step by step, so you can check every move. If you can’t audit it, you can’t trust it.
Plans
Self-hosting is free, always.
| Plan | What you get | Price | Status |
|---|---|---|---|
| Brahma | Run it yourself on your own machines. Share a node with your team, signing in with passkeys, Google or Microsoft. | $0, forever | Private beta |
| Pro | Chat on the hosted mesh, with higher limits and the explorer built in. | At launch | Coming soon |
| API | Pay per token on the hosted mesh. The same open-weight models, no servers to run. | At launch | Coming soon |
| Enterprise | Private meshes on your hardware, SAML and Verifiable Credentials sign-in, support, and a commercial license. | Talk to us | Contact us |
Benchmarks
How fast it runs, and how it grows.
Measured on laptops, phones and cloud CPUs. No GPUs.
From 24 machines working as one
Granite 3.1 with four chats on every machine. Each chat on the 24-machine mesh ran at 6.6 to 6.9 tokens/s.
Granite 3.1 1B-A400M on Apple Silicon. First token in about 0.16 s once warm.
Qwen3-235B-A22B across 9 ordinary CPU machines, all 12,032 experts placed, at 0.4 tokens/s.
One chat on a mesh spread across seven AWS regions, with tuned settings. With five Pixel phones joined over the internet, a chat ran at 5.6 tokens/s against 5.7 without them.
Roadmap
What we're building next.
Bigger models
- Qwen3.6-35B-A3B, then DeepSeek-V4-Flash (291B) and Kimi K2.6 (1 trillion parameters)Planned
- Frontier models: Qwen3.8, DeepSeek-V4-Pro and Kimi K3 on a mesh of cloud machinesPlanned
- Z.ai's GLM models, from laptop size to GLM-5.3 (753B)Planned
- Works with apps built for Ollama and imports your Ollama modelsPlanned
Your machines
- Two-Mac meshes, a Mac with a Raspberry Pi, and physical Pi boardsNext
- The Android app on a Samsung Galaxy S26 and a tablet; the iPhone and iPad app compiledNext
- Faster on Apple Silicon, with an MLX backendPlanned
- A browser tab as an expert host, with WebGPU where availablePlanned
Hosted Brahma
- The mothership at mybrahma.com: chat and a pay-per-token APIPlanned
- A consumer app, and private meshes for enterprisesPlanned
- Payouts to node operators for verified workPlanned
- A public mesh that machines behind home routers can join directlyDesign
Engineering
Four layers. One model spread over many machines.
Vertices, not matrices
Tokens, attention heads, neuron groups and experts become vertices that can live on different machines.
Machines that sort themselves
Nodes find each other, size themselves, and move experts as machines come and go.
Faster repeats
Speculative decoding speeds up replies that reuse text from the prompt, with the same output.
A Rust data plane
An optional native path carries calls between hosts; a node keeps its own experts in-process.
Why devices
The current plan for AI is to build ever larger GPU datacenters. Brahma takes the other path: useful AI built from the machines people already own.
They draw a lot of power
Large GPU clusters need dedicated power and cooling, and each generation asks for more. The cost lands on grids, water supplies and the climate.
They concentrate control
When only a few companies can afford the hardware, they decide who gets access, at what price, and what the models will and won't do.
Our machines sit idle
Laptops, desktops, home servers and spare cloud instances spend most of the day waiting. Together they are a huge pool of compute nobody is using.
Who it's for
Built for people who want their own AI.
Hobbyists and homelabs
Turn the laptop, the old desktop and the Pi on the shelf into one model server you understand end to end.
People who want privacy
Prompts and answers stay on machines you control. Nothing has to leave your network.
Researchers
Every head, expert and neuron group is a vertex you can watch, query and move.
Businesses
A private mesh on devices you already have, with support and a commercial license when AGPL doesn't fit.
For companies
A private mesh on machines you already own.
Run Brahma on your company's laptops, desktops and servers. Prompts stay inside your network, and modest inference needs no new GPUs. A commercial license covers embedding, hosting, and teams whose policy rules out AGPL.
Open source
Open source, with a way to sustain it.
License
Brahma is AGPL-3.0. Run a modified version as a service and you share your changes. A commercial license is there for teams that can't.
Contribute
Contributions are welcome when the repository opens. Contributors sign a CLA, which keeps the commercial license that funds the project possible.
Sponsor
Sponsorship pays for development and test hardware. Get in touch if you or your company want to help.
Earn for spare compute
Node operators will contribute spare capacity to the hosted mesh and earn for verified work. The accounting is built: signed receipts of every job. Coming soon
Bring a device. Join the mesh.
Invites go out in the order people joined, a batch at a time. Run a node on your own Mac or Linux machine, add a second one, and tell me what breaks.
Business, research or press? hello@asyml8.com