Multi-LLM Mesh

How DOF-MESH coordinates heterogeneous LLM nodes — cloud API and local — into a single governed inference layer.

The mesh is a coordinated network of LLM nodes. DOF routes tasks to the optimal node based on task type, availability, cost, and latency — with TTL backoff, circuit breaking, and threshold consensus for critical decisions.

---

Cloud API Nodes

Node ID	Provider	Speed	Cost	Best for
minimax	MiniMax M2.1	—	1,000 req/day free	Primary provider
cerebras-llama	Cerebras	868 tok/s	Free	Fast inference
deepseek-coder	DeepSeek	120 tok/s	$0.27/M	Code generation
sambanova	SambaNova	200 tok/s	Free	Large context
nvidia-nim	NVIDIA NIM	150 tok/s	Free tier	Complex reasoning
glm-5	Zhipu GLM	100 tok/s	Free	Multilingual
gemini-flash	Gemini	—	Free/web	Analysis
groq	Groq	—	Free tier	Low-latency

Local Nodes (Ollama)

Node ID	Base model	Role	Context	Size	Cost
local-agi-m4max	qwen2.5-coder:14b	Autonomous local AGI	65K	9.0 GB	$0.00
dof-coder	qwen2.5-coder:14b	Code implementation	65K	9.0 GB	$0.00
dof-analyst	qwen2.5-coder:14b	Data analysis	65K	9.0 GB	$0.00
dof-guardian	qwen2.5-coder:14b	Security (Cerberus Prime)	65K	9.0 GB	$0.00
dof-reasoner	deepseek-r1:14b	Strategic planning	65K	9.0 GB	$0.00
dof-voice	qwen2.5-coder:14b	Voice — long context Spanish	65K	9.0 GB	$0.00
dof-voice-fast	gemma2:9b	Voice — realtime Spanish	8K	5.4 GB	$0.00

The alias local-qwen no longer exists. Use local-agi-m4max for local sovereign compute. Set LOCAL_MODEL_ENDPOINT in .env to your Ollama server URL.

Local Node System Prompts

local-agi-m4max: "You are local-agi-m4max — the autonomous local AGI node of the DOF Mesh"
dof-coder:       "You are DOF-CODER — code intelligence node of the DOF Mesh"
dof-analyst:     "You are DOF-ANALYST — the data intelligence node of the DOF Mesh"
dof-guardian:    "You are CERBERUS PRIME — Active Defense Commander of the DOF Mesh"
dof-reasoner:    "You are DOF-REASONER — strategic intelligence, orchestration, and planning"
dof-voice:       "Eres la VOZ del DOF Mesh Legion. Tu creador es Cyber Paisa"
dof-voice-fast:  "Eres la voz del DOF Mesh Legion"

---

Starting the Mesh

python3 core/api_node_runner.py \
  --nodes deepseek-coder cerebras-llama gemini-flash local-agi-m4max \
  --daemon

---

TTL Backoff

When a provider fails or rate-limits:

1st failure → 5 min cooldown
2nd failure → 10 min cooldown
3rd failure → 20 min cooldown

The circuit breaker transitions: CLOSED → OPEN → HALF_OPEN → CLOSED

---

Threshold Consensus

For critical decisions, DOF requires N-of-M agreement across mesh nodes:

from core.threshold_consensus import ThresholdConsensus

consensus = ThresholdConsensus(threshold=3, total=5)
result = consensus.vote(action="transfer_1000_USDC", voters=selected_nodes)
# {"approved": True, "votes": {"approved": 4, "rejected": 1}}

---

Key Modules

Module	Responsibility
core/mesh_router_v2.py	Smart routing by task type
core/mesh_scheduler.py	Priority queue (CRITICAL > HIGH > NORMAL > LOW)
core/mesh_load_balancer.py	Distribution across equivalent nodes
core/mesh_circuit_breaker.py	CLOSED / OPEN / HALF_OPEN per provider
core/mesh_orchestrator.py	Task routing + horizontal scaling
core/node_mesh.py	NodeRegistry + MessageBus
core/threshold_consensus.py	N-of-M voting for critical decisions
core/web_bridge.py	Playwright — browser-accessible LLMs without API

---

Running the Mesh

Launch the full mesh with Docker or CLI

Providers & Agents

Configure provider chains and 17 agents