§ 03 · Technical Specifications
The numbers, exactly.
Implementation-grade specifications covering formats, quantization, hardware envelopes, compatibility, and failure semantics. Engineers can build against this page alone.
3.1
Supported model formats
3.2
Quantization matrix
Quantization is selected per-device at model download time based on the capability probe. Nodes can hold multiple quants and the Router picks at dispatch.
3.3
Browser compatibility matrix
3.4
Hardware envelopes
Desktop (native client)
| Minimum | 4-core x64/arm64 · 8GB RAM · integrated GPU · 10GB free disk |
| Recommended | 8-core · 16GB RAM · 6GB VRAM discrete GPU or Apple M1+ |
| Premium | 16-core · 32GB RAM · 12GB+ VRAM · runs 7B Q5 @ 45 tok/s |
Mobile
| iOS minimum | iPhone 13 / iPad M1 · iOS 17 · 4GB RAM |
| iOS recommended | iPhone 15 Pro+ · A17 Pro / M-series · 8GB+ RAM |
| Android minimum | SD 8 Gen 2 or equivalent · 8GB RAM · Android 14 |
| Android recommended | SD 8 Gen 3, Dimensity 9300+ · 12GB RAM · Android 15 |
3.5
Security model
| Transport | TLS 1.3 only. HSTS preload. ECDHE + ChaCha20-Poly1305 preferred. |
| Task envelopes | AES-256-GCM, keys per-session, derived via HKDF from TLS exporter |
| Model integrity | Every model bundle signed Ed25519. Refused on signature mismatch. |
| Sandbox (browser) | Workers run in isolated realms. No DOM access. No network beyond signed Coordinator URLs. |
| Sandbox (desktop) | Rust worker in seccomp/App Sandbox. No FS outside cache dir. |
| Prompt exposure | Local routes: prompt never leaves device. Mesh routes: encrypted end-to-end to peer; peer runs in isolated worker; result encrypted back. |
| Peer anonymity | Requesting node never learns peer identity; Coordinator acts as blinded relay. |
| Data retention | Coordinator retains metadata only (no bodies) for 30 days. Configurable to 0. |
| Compliance | SOC 2 Type II (target Q3), GDPR DPA, HIPAA BAA (Enterprise), ISO 27001 (target 2027) |
3.6
Task verification
Peer-dispatched tasks are verifiable without leaking plaintext:
Peer-dispatched tasks are verifiable without leaking plaintext. A configurable sample of tasks is duplicated to independent verifier nodes. Outputs are compared under deterministic decoding conditions to detect discrepancies. A hash mismatch escalates both outputs to a third verifier; the minority output is flagged suspicious and the node's reputation is penalized. Repeated suspicious results within a rolling window trigger node quarantine.
3.7
Failure recovery & retries
3.8
Rate limiting & throttling
| Per API key | 600 req/min sliding window · 100k tokens/min |
| Per end-user | 60 req/min · customer-scoped (enforced in SDK + coordinator) |
| Per node (mesh work) | 120 tasks/min · prevents peer exploitation |
| Burst allowance | 2× limit for 10s, token-bucket refill |
| 429 response | includes Retry-After header + X-MeshInfer.AI-Reset-At epoch ms |