Start here
NeuroShard is building a living, decentralized AI: a shared model executed across contributed machines and improved through evaluated learning. The public code provides its native ledger, sharded execution, work verification, quality admission, and payment mechanisms.
The website presents the vision. These docs explain what exists, how the demonstrations work, and how to participate in the published execution profiles.
What is demonstrated
| Area | Measured evidence | Scope |
|---|---|---|
| Useful learning | Three admitted specialist cohorts improve measured single and combined answers and preserve tested earlier answers. | Bounded source-backed knowledge tasks; general ChatGPT-level capability is not established. |
| Automatic continuation | The latest two useful cohorts run consecutively through data admission, training, audit, and promotion. | Approved immutable feeds and one operator; arbitrary public data and indefinite learning remain unproved. |
| Distributed serving | Seven GPU hosts served the accepted graph; no physical host contained the whole backbone. Paid streaming and provider-process recovery passed the frozen gate. | Two simultaneous requests, 64 output tokens, public data; GPU service retired September 20, 2026. |
| Native lifecycle | Training rewards, separate quality promotion, inference payment, and refunds have reproducible execution and ledger evidence. | Full numerical replay is expensive; finite sponsorship is not a self-sustaining market. |
The fixed checklist records 5/6 complete against bounded demonstration criteria. Item 4 remains open for dependable permissionless hosting and independent operation. This is not a percentage measure of progress toward a general, continually improving global assistant.
Choose the right network
Growing-model research alpha
The GPU alpha closed on September 20, 2026 at the owner's request. The historical guide pins its source, genesis, public descriptor and observer. Its starter-credit and chat commands no longer provide a live GPU service.
- New admissions are closed and all seven GPU hosts are retired.
- Ledger availability is scheduled through September 26, 2026, 23:49 UTC for expiry and refunds.
- Model artifacts, execution evidence and private recovery records were preserved.
- This serving window ran no new learning cohorts.
- Prompts, prior context, and settled replies are public ledger data.
This page documents a finite experiment, not an always-on service guarantee. Use the source and evidence for reproduction. No replacement GPU deployment has been announced. Ledger observation remains available during its separate retention window.
Earlier 0.4.0 CPU baseline
The baseline guide, model card, protocol, and API describe the older SmolLM2-135M-Instruct adapter network. The PyPI 0.4.0 client and website browser chat use this network.
Check its live status before submitting work or payments. It was observed stalled on September 20, 2026; the website displays the gateway’s current state. Its balance and release instructions are separate from the growing-model alpha.
Understand the architecture
- Scaling design: compute groups, resource-backed growth, quality, availability, and what added machines can actually provide.
- Complete answering-system admission: bind routing, planning, model weights, tokenizer, and generation to the evaluated serving graph. This document is an RFC; use linked implementation and result records to determine activation status.
- Provider runtime and hosted chat: discovery, execution, conversation identity, and payment recovery.
- Finite sponsorship and funded auditing: what verification costs and which work is actually funded.
More hardware can add memory, throughput, or redundancy. Model improvement requires useful data, successful learning, and evaluation; lower response latency depends on the workload and the links between machines.
Build or contribute
The proposed local AI appliance and business offering are described in product direction. A supported private installation, arbitrary hardware compatibility, and separation of private business workloads from public contributions are not yet shipped.
For software work, use the contribution guide and research roadmap. Teams and hardware partners can contact info@neuroshard.com with a non-confidential description of what they want to build.