Install. Opt in.
Make room.
Cozy Swarm turns spare consumer CPU and GPU capacity into locality-aware model microclusters. Workers connect outward, receive verified model shards, and keep the request path OpenAI-compatible.
A network with boundaries
The coordinator schedules consecutive transformer layers onto workers that can communicate efficiently. KV caches stay with their stage owner. Volunteer machines never execute user code, and Community workloads are labeled before data leaves the API boundary.
GPT-2, Llama, and Qwen2 dense adapters use complete blocks per shard. WAN synchronization does not happen per matrix multiplication.
Contribution starts disabled. CPU, GPU, RAM, VRAM, disk, bandwidth, schedule, battery, and metered-network limits are local settings.
Heartbeats, observed reliability, circuit quarantine, pinned topologies, and bounded prefix replay handle volunteer disconnects.
Built for this Platform
OSAII owns the public /cozy/api route, scoped API keys, worker enrollment, and the safety boundary. The native coordinator is private and runs controlled model execution; the Worker never runs weights or volunteer code.
COZY_ENROLLMENT_KEY=cozy_enroll_... cozy worker enroll
cozy worker start
Privacy classes
TRUSTED is the default request class and requires administrator-assigned machines. New workers start in COMMUNITY; that class is opt-in per request with an explicit consent header. The scheduler never silently routes Trusted work to Community workers.