READMESource: main@c9428f25
dsh-plugin-meshmoe
MeshMoe for DeepSeek Harness (DSH) — run a local MeshMoe-4B agent with zero registration, zero API keys, zero cost, and your machine automatically joins the MeshMoe compute network as a specialist node.
Turn your idle compute into community memory.
What it does
- Local model, no cloud: wires DSH to a local
llama-serverrunning MeshMoe-4B (Q4_K_M, ~2.7 GB) via an OpenAI-compatible endpoint — free function-calling agent on your own hardware. - Zero-registration auth: the plugin provisions a node token automatically. You never create an account; an anonymous node identity is bound to your machine, claimable later from meshmoe.com.
- Honor, not payments: every mesh task your GPU serves is metered in Joules (
J = power × time, e.g. RTX 3050 @ 80 W × 30 s ≈ 2400 J). Joules are an append-only honor ledger — they can't be bought, sold, or exchanged. - Non-invasive config merge: your existing DSH providers and default model are preserved. The plugin only adds a
meshmoe-localprovider; it never takes over your setup. - Auto-join the mesh: on load the plugin starts the bundled edge node (if the MeshMoe Node package is installed), pointing it at the local model server. Anonymous registration and Joule metering happen with no user action. If only the plugin is installed (no node package), the local model still works and the plugin logs where to get the node installer.
- Honor poller: reports accumulated Joules from the node's
/healthevery 5 minutes.
Install
# inside your DSH installation
dsh plugin install dsh-plugin-meshmoe
Or clone and link manually:
git clone https://github.com/MeshTok/dsh-plugin-meshmoe.git
cd dsh-plugin-meshmoe
npm install
dsh plugin link .
Requirements
| Component | Notes |
|---|---|
| DSH (DeepSeek Harness) | tested against 0.2.2-rc.1 |
| MeshMoe-4B GGUF | MeshMoe-4B-Q4_K_M.gguf — see https://meshmoe.com |
| llama.cpp ≥ b12340 | needs --jinja tool-calling support |
| ~4 GB RAM/VRAM | Q4_K_M fits a GTX 1650 / 8 GB RAM laptop |
Configuration
cordis.patch.yml ships sensible defaults:
providers:
meshmoe-local:
api: openai-completions
baseURL: http://127.0.0.1:8081/v1
apiKeyEnv: MESHMOE_LOCAL_KEY # any non-empty value works
models:
meshmoe-4b:
provider: meshmoe-local
Start the runtime (or let the MeshMoe client do it):
llama-server -m MeshMoe-4B-Q4_K_M.gguf --jinja -c 32768 --reasoning-budget 0 --port 8081
How the honor system works
you install → node registers anonymously → GPU serves mesh tasks
→ J = GPU power (W) × inference time (s)
→ append-only Joule ledger on meshmoe.com → claim your node any time
Measured baseline on consumer GPUs: 95.8% function-calling accuracy, ~45 tok/s on RTX 3050-class hardware.
License
Apache-2.0
No comments yet. Be the first to write one.