Distribute the model
Execute model shards across contributed machines. Added hardware can provide memory, serving capacity, and redundancy.
The open technical foundation for a shared model, contributed compute, and verified learning.
Execute model shards across contributed machines. Added hardware can provide memory, serving capacity, and redundancy.
Admit useful learning only after measured improvement and preservation of earlier answers. Inspect successful cohorts and failed experiments.
Use native consensus for work commitments, verification, rewards, quality promotion, and paid inference.
| You want to… | Begin with |
|---|---|
| Understand what exists today | Current capabilities and limits and the fixed checklist |
| Explore the vision for your team or hardware | Product direction |
| Try the growing-model research alpha | Source-pinned alpha joining guide |
| Inspect learning and serving results | Continual admission and operated alpha |
| Run the earlier 0.4.0 CPU baseline | Baseline guide and model card |
| Build on the protocol | Architecture, all guides, and contributing |
The growing-model alpha and the earlier 0.4.0 baseline are separate networks. The current alpha has a finite admission window ending September 22, 2026 at 11:09 UTC, subject to earlier funding limits. Its accepted weights remain fixed during that window; it does not run new training. Read start here before choosing a client or sending requests.
The published demonstrations use infrastructure controlled by one operator. Independent permissionless operation and broad assistant quality remain open work. Public prompts and responses are ledger data; trial balances have no redemption promise.
These pages are generated from the public protocol repository. The source manifest records the exact revision and source hash for every generated guide. Historical research is labeled separately.