02 · TECHNOLOGY
UNDER THE HOOD
MESH routes your AI workloads to the lowest-latency, highest-throughput nodes automatically. The protocol is designed for adversarial environments — every result is consensus-validated before delivery.
REQUEST PATH — HOW YOUR JOB RUNS
1. SUBMIT — API call or SDK. Job enters the MESH queue (Redis sorted-set).
2. MATCH — WorkMatcher pairs job to the best available node in <100ms.
3. ASSIGN — Node receives assignment via WebSocket push (or HTTP poll fallback).
4. EXECUTE — ONNX Runtime / WebGPU runs inference in the browser sandbox.
5. VALIDATE — Result hashes replicated across N nodes. Consensus threshold checked.
6. DELIVER — Verified output returned. Job marked complete. Billing settled.
PROTOCOL GUARANTEES
| Delivery SLA | 99.7% |
| Result validation | Hash-based consensus (configurable replication) |
| Fault tolerance | Automatic lease re-assignment on node drop (<15s) |
| Nonce proof | Per-job nonces prevent replay attacks from adversarial nodes |
| Epoch locking | Lease epochs prevent stale result injection |
| Transport | WebSocket push / HTTP poll fallback |
| Node selection | Latency + VRAM + utilization scoring |
COMPUTE CAPABILITIES & WORKLOADS
| Generative AI | Multi-modal synthesis (Text/Image/Video/3D). Examples: FLUX.2, LTX-2.3, Hunyuan3D. |
| Scientific Computing | Monte Carlo simulations, molecular dynamics, mathematical modeling via WebGPU. |
| Data Engineering | Vector processing, point cloud transformations, large-scale feature extraction. |
| Media Pipeline | Real-time 4K/8K upscaling, depth estimation, background removal, frame interpolation. |
| Custom Workloads | Deploy any ONNX or WGSL-based compute kernel across the distributed grid. |
EXECUTION MODES
SINGLE
One node processes the entire job. Ideal for small models (BK-SDM-Tiny, BiRefNet). Lowest coordination overhead.
PARALLEL (TILED)
Input split into tiles, each assigned to a different node. Results merged. Used by ESRGAN and SeedVR2 upscalers for large inputs.
SHARDED
Model split across N nodes — each node holds a subset of weights. Nodes communicate via WebRTC P2P. Used by FLUX.2-klein, LTX-2, Wan 2.2, Hunyuan3D 2.1, TRELLIS.2-4B, ACE-Step.
Your browser becomes a compute node. ONNX Runtime Web runs inference directly on your GPU via WebGPU — no driver install, no CLI, no server.
NODE LIFECYCLE
1. CONNECT — Node registers via MESH SDK, declares VRAM + compute capacity.
2. ADVERTISE — Sends
node-available message over WebSocket. Enters Redis availability pool. 3. RECEIVE — WorkMatcher pushes assignment. Node holds lease (Redis TTL).
4. EXECUTE — Inference runs in Web Worker via ONNX Runtime. GPU stays dedicated to the job.
5. SUBMIT — Result hash + output uploaded. Lease released.
6. READY — Node re-advertises. Heartbeat refreshes Redis TTL every 30s.
SYSTEM REQUIREMENTS
| Browser | Chrome 113+ / Edge 113+ (WebGPU required) |
| GPU | Any WebGPU-capable GPU (NVIDIA, AMD, Apple Silicon, Intel Arc) |
| Min VRAM | 4 GB (Tier-3). 8 GB (Tier-2). 16 GB+ (Tier-1) |
| Bandwidth | 10 Mbps+ recommended for model P2P transfer |
| Storage | OPFS cache — 2–20 GB per model set (browser file system) |
| Uptime | No minimum — a device is usable whenever it is online |
NODE TIERS
| Tier 1 | 16 GB+ VRAM | PREMIUM JOBS · 2.2× RATE |
| Tier 2 | 8–15 GB VRAM | STANDARD JOBS · 1.4× RATE |
| Tier 3 | 4–7 GB VRAM | LIGHT JOBS · BASE RATE |