Join a network
Run hetero-worker --join CODE on each machine. It reports its cores, memory and GPUs, and receives a revocable credential.
Distributed compute · beta
Run models larger than
any one of your machines.
Hostwatch connects your Macs, PCs and servers into compute networks. weavatrix-hetero spreads a model’s layers over the memory they lend — a graphics card, Apple silicon, or plain RAM — and answers exactly as a single machine would.
Set its share
Choose how much of each machine the network may use, when, and for what. Change it any time from the web or the app.
Plan the model
The coordinator probes the machines that are free, measures them, and splits the layers by memory first and speed second.
Run it exactly
Stages compute their layers in turn; the coordinator drafts and verifies. Clients see one model behind OpenAI’s or Ollama’s API.
Connect any machine
A Mac lends its GPU through Metal, a PC its graphics card through Vulkan or DX12, any machine its processor and memory. One command with a one-use code, and it appears in your networks.
Decide what each one lends
A power share, memory per GPU or RAM, the hours it works in its own time, the kinds of jobs it takes, how many at once. Pause it from the web or your iPhone; it follows within seconds.
Model size grows with machines
Each machine measures what it can hold and how fast it runs a layer. The plan puts the model’s layers where they fit and run fastest, so a model none of them can load runs across them.
The same answer as one machine
Every token is the model’s own greedy choice. Proposals from drafts and lookups are verified, never trusted, and a run is reproducible to the bit for the same placement.
Networks and keys
Separate machines by purpose. A network key lets the hetero CLI and its OpenAI- and Ollama-compatible server place models across that network’s machines, and nothing else.
Community pool
Offer spare power to other organizations and use theirs, within the share you set. Free during the beta.
Measured, not projected
What it has done so far.
Runs recorded in the weavatrix-hetero repository, on a Mac and Linux machines in containers with hard memory limits.
Qwen3-4B’s layers in BF16 ran across two Linux machines of 4.5 GiB each. Alone, each refused before loading; a naive load was killed by the kernel.
63 of 63 logits rows bit-identical to the same layout on one machineThat model answered across the two 2-vCPU machines at 10.6 tokens per second, the Mac coordinating and drafting with Qwen3-0.6B.
results/distributed-two-linux-20261006Qwen3-0.6B gave the same text on Apple’s GPU, WebGPU, CPUs on macOS and Linux, Vulkan, and split across a Mac and a Linux machine joined through a Hostwatch network.
results/compute-network-20261007Beta.
Compute networks and the community pool are free while in beta. The job queue runs through Hostwatch; the computation itself flows directly between your machines, so they must reach each other on one network, over a VPN such as WireGuard or Tailscale, or through open ports. A relay through Hostwatch is planned. What passes where →
Probes, plans, stages, verification, and what travels through Hostwatch.
Connect a machineBuild the worker for macOS, Linux or Windows and join a network with a code.
Control and permissionsPower share, memory, hours, job kinds, roles and the audit trail.
CLI and APIsNetwork keys, hetero run --across all, and the OpenAI and Ollama APIs.
Community poolLend spare power to other organizations, and use theirs. Free in beta.
Security and privacyCredentials, isolation, what each party can see, and how to revoke.