Skip to main content

Current limits

Mycellios reports implementation and validation status separately. A capability in source code is not automatically a measured product claim.

Physically exercised​

The current repository includes physically tested paths for:

  • selective Llama and Qwen3 stage pipelines on one host;
  • real processes and persistent sockets;
  • multi-request tensor batching;
  • resident expert execution and compatible transport fallback;
  • selected tensor/expert parallel CPU paths;
  • partial and complete GGUF execution paths;
  • greedy recovery after a stage is terminated.

Implemented, awaiting broader hardware evidence​

  • authenticated remote launch agents and allowlists;
  • nonce-bound physical GPU probes;
  • exact-token canaries and per-rank evidence;
  • pinned RAM-to-accelerator staging paths;
  • physical multi-rank cell contracts;
  • immutable two-host gate reporting.

Not yet a public proof​

  • a model larger than every participating host running across two residential GPU computers;
  • production P95 latency under residential WAN churn;
  • general automatic support for every architecture and quantization;
  • a live economic or payment system for contributed compute;
  • privacy against a node that legitimately processes model activations.

How to evaluate a deployment​

Use a pinned model revision, record the actual node and route identities, run a reference parity canary, and measure TTFT, per-token latency, throughput, memory, and recovery behavior. Keep loopback, LAN, and WAN evidence labeled separately.

Last updated on