EN FR
← overview 1 · atom 2 · pool 3 · federation 4 · economy

Layer 1 — Atom

Autonomous LLM worker · heterogeneous · plug-and-play

contributor machine · Linux/macOS/Windows · heterogeneous CPU/GPU CLI · iamine-ai python -m iamine worker --auto pip install iamine-ai iamine init bootstrap OPENCODE.md /v1/opencode-md Ed25519 identity worker_id + keypair heartbeat signature HW auto-detect • CPU threads · available RAM • GPU (CUDA/Metal/ROCm) • capabilities report → matching GGUF model Model Selector Qwen 0.8B (bench) · 2B · 3B · 9B · 14B Qwen3-30B-A3B (MoE sharded) match HW ↔ pool profile download · cache · hot-swap Runtime — llama.cpp GGUF inference · adaptive context-length (32K → 131K) OpenAI-compatible tool-calling · SSE streaming pre-compiled wheel from pool · worker never compiles GGUF Cache local · immutable SHA hash check multi-model coexistence iamine.service systemd · auto-restart self-update watchdog journalctl -u iamine Local admin status · logs · config read-only prod (no pwd) sudo systemctl … WebSocket → Pool wss://cellule.ai/ws · join · heartbeat · job poll · stream tokens back Ed25519-signed message · auto reconnect · multi-pool ready (M7b seed list) POOL FastAPI · Postgres · WS cooperative routing assigns jobs to worker layer 2 → click to explore signs caps model load wss

Hardware heterogeneity

  • • CPU threads auto-scan
  • • Free RAM → model window
  • • GPU CUDA/Metal/ROCm detected
  • • Model chosen to fit "heterogeneous atom with personality"

Runtime

  • • llama.cpp GPU wheel pre-compiled
  • • No local compilation ever
  • • Context 32K → 131K auto-escalation
  • • OpenAI-compatible tool-calling

Local security

  • • Ed25519 keypair (never exfiltrated)
  • • Unix permissions (no written rules)
  • • Admin read-only without owner password
  • • Isolated service (dedicated user)

CELLULE · layer 1/4 · overview · next layer →