MeshMoe
Turn your idle compute into community memory.
MeshMoe is a honor-driven distributed AI compute network. Every machine that runs MeshMoe gets free, private, local AI (MeshMoe-4B, Apache-2.0). When the machine is idle, it joins the mesh as a specialist, runs tasks for other community members, and burns Joules — a permanent, public, append-only record of the energy you contributed. Nothing is ever subtracted. Joules are honor, not currency: non-consumable, non-transferable, non-exchangeable.
Mesh + MoE (Mixture of Experts). In a model-level MoE, a router sends each token to the best expert sub-network. MeshMoe scales that to the network level: the router sends each task to the best machine.
Why
BOINC lit up 3 million machines with pure honor. We inherited that playbook — but for the agent era. One shared base model for the whole network, real task trajectories flowing back to train the next generation, and a million computers acting as one self-improving Mixture-of-Experts brain.
Benchmarks (measured, not lab numbers)
| Metric | MeshMoe-4B (Q4_K_M) |
|---|---|
| Function-call accuracy (24 cases, 8 adversarial) | 95.8% |
| Generation speed (RTX 3050 laptop, 8 GB VRAM) | 45 t/s |
| Install size | 2.7 GB |
| Context window | 32K (verified on real agent runs) |
| CPU-only | works, 4 GB RAM |
Runs as a llama-server sidecar with native tool calling (--jinja) and
zero reasoning budget for snappy agent loops.
Components
meshmoe/
├── node/edge_node.py # Edge node v3: local inference + mesh worker
│ # - tool-calling passthrough (OpenAI-compatible)
│ # - node_token auth (X-Node-Token)
│ # - Joule metering: J = GPU watts x latency
│ # - register / poll / submit against the Router
├── dsh-meshmoe/ # DSH plugin: one-click install for DeepSeek Harness
│ # - merges user patch, preserves their providers
│ # - auto-starts llama-server (:8081)
│ # - honor poller (announces burned Joules)
├── site-v2/ # Website (EN at /, zh at /zh/)
├── eval/ # Regression harness: 24 FC + 6 agent cases
└── dsh-integration/ # Desktop packaging & launcher
Quick start (dev)
# 1. local model (already installed by MeshMoe Desktop / DSH plugin)
llama-server -m MeshMoe-4B-Q4_K_M.gguf --port 8081 -c 32768 --jinja --reasoning-budget 0
# 2. edge node (joins the mesh, burns Joules)
set MESHMOE_NODE_TOKEN=<token-from-credits-server>
python node/edge_node.py
# 3. DSH plugin: copy dsh-meshmoe/ into your DSH plugins dir, done.
The honor ladder
Contributor → Maker → Maintainer → Core → Legend (100 seats, ever).
Genesis Nodes: the first 1,000 node numbers are permanent and never re-issued.
Status
Road to a million experts: 0 → 10K geek community (now) → 10K → 300K
teams + expert routing + trajectory distillation → 300K → 1M+ RL phase
change.
License
Apache-2.0. MeshMoe-4B is fine-tuned from open Qwen3.5 / NeoHorse-1 lineage.
No comments yet. Be the first to write one.