Shard
Neural Mesh
unreachableBootstrapping
Neural Gateway
default-modelv0.0.0
SYNCING MESH
0 NODES · 100% RELIABILITY
Press Enter to send, Shift+Enter for new line
Live Network Status
Network data unavailable — single-node mode active.
Traditional Cloud AI vs. Shard Network
💰 Cost$0.002–$0.06 / 1K tokensFree (compute-for-access)
🔒 PrivacyYour data on someone else's serverLocalhost-first routing
📈 ScalabilityBuy more GPUsMore users = more GPUs
🛡️ ResilienceSingle point of failureSelf-healing P2P mesh
⚡ LatencyNetwork RTT + queue waitLocal draft + network verification
🔌 APIProprietaryOpenAI-compatible drop-in
How Contribution Works
Scouts draft likely next tokens in-browser, shard nodes verify them in parallel, and clients receive trusted output with lower latency.
Interactive Inference Demo
Output will stream here…
0Draft tokens contributed by Scouts
0Verified by Shard node
Built With
- GGUF Runtime: runs quantized model weights efficiently so verifier nodes can serve high-quality output with lower memory pressure.
- libp2p: provides peer discovery and resilient transport so nodes can coordinate over a decentralized mesh.
- WebLLM: powers browser-based draft generation so contributors can help without installing native runtimes.
- WebGPU: lets modern browsers execute local draft inference quickly with GPU acceleration.
- Rust: drives the daemon and networking layer for deterministic performance and safety under load.
- FastAPI: supplies OpenAI-compatible HTTP endpoints and orchestration hooks for app integrations.
Project Signals
Benchmark callout: documented API performance sample shows 850 tokens/sec distributed throughput and 245 req/min in live cluster mode, compared with a nominal single-node baseline of ~320 tokens/sec (~2.6x uplift).