Run the open-source client, share your GPU, earn 70% of every edge-served request. Open client (MIT) — audit before running.
option A · windows desktop app (recommended)
Chat with 100+ open models (cloud + edge), or run models locally — free & private.
Share your GPU as a node in one click: pick a model → download → run. Earn 70% of every edge-served request.
Windows 10/11 x64 · uses system WebView2 (auto-fetched if missing) · unsigned — click "more info → run anyway" on SmartScreen.
macOS / Linux builds coming.
option B · linux / macos / datacenter (Python node client)
# 1. Install Python 3.10+ if you don't have it$ python3 --version
# 2. Clone the open-source client (MIT — audit before running)$ git clone https://github.com/OpenMeshMoE/MeshMoE.git
$ cd MeshMoE
# 3. Install dependencies (llama-cpp-python for local inference)$ pip install -r requirements.txt
# 4. Link to your account (get a key at meshmoe.com/app/keys — earnings land in your balance)$ export MESHMOE_API_KEY=moe-xxxxxxxxxxxx
# 5. Pick your expert model and run$ python edge_node.py
[meshmoe] registered as edge-node-xxxx (tier: standard)
[meshmoe] model loaded (Q4_K_M)
[meshmoe] polling router for tasks...
[meshmoe] task #482: completed in 1.4s · +3 credits# ── datacenter / existing vLLM · sglang · TGI · llama-server ──
# Point at any OpenAI-compatible endpoint you already run — no model download:$ export MESHMOE_INFER_URL=http://10.0.0.5:8000
$ export MESHMOE_INFER_MODEL=deepseek-v3.2
$ export MESHMOE_MODEL=DeepSeek-V3.2 MESHMOE_TIER=heavy
$ python edge_node.py
option C · windows (Python, from source)
> git clone https://github.com/OpenMeshMoE/MeshMoE.git
> cd MeshMoE
> pip install -r requirements.txt
> set MESHMOE_API_KEY=moe-xxxxxxxxxxxx
> python edge_node.py
choose your expert model
tier
model
vram needed
earn rate
good for
light
GLM-4-9B
12 GB
70%
general chat, fast
standard ⭐
DeepSeek-R1-Distill-Qwen-14B
16 GB
70%
code, math, reasoning — main network workhorse
standard
Qwen3-Coder-14B
16 GB
70%
pure code completion
heavy
DeepSeek-R1-Distill-Qwen-32B
24 GB+
70%
heavy reasoning
how earnings work
When your node serves a request:
user pays N credits (70% of cloud price — edge discount)
├─ 70% → your account (the node owner)
└─ 30% → meshmoe (network ops, router, infra)
No idle payout — only real edge-served requests earn. Credits are spendable on the network (not withdrawable for cash).
Cloud-fallback requests (when no edge node serves) earn nothing.
audit the client (open source, MIT)
Before running, you should verify the client:
✓ Does not mine cryptocurrency
✓ Does not exfiltrate your files or personal data
✓ Only runs the model you explicitly selected
✓ Only responds to router-dispatched inference tasks
Future — Personal Expert Models:
Have private domain data (legal cases, medical records, internal docs)?
A future client release will let you fine-tune your own expert (R1-Distill + LoRA)
and contribute it as a specialized node. Your data never leaves your machine;
your expertise earns credits when others query that domain.
Read the vision →