Open Weights Network - Inference Mesh PoC

This repository is the inference-first track for distributed open-weight model serving.

OWN = Open Weights Network

Goal:

make open/open-weight models usable through a people-powered inference mesh before attempting distributed training

Root design choice

Option A

open weights = download a model and run it alone

Why not: local-only execution does not create shared access, routing, availability, or a public alternative to closed APIs.

Option B

open weights network = one central hosted API for open models

Why not: a central API recreates the corporate dependency layer.

Chosen direction

open weights network = decentralized inference mesh with verifiable routing, node reputation, privacy levels, and OpenAI-compatible access

PoC loop

InferenceTask → NodeRegistry → Router → Health/Privacy/Capability gates → chosen NodeCard → InferenceReply

Quick start

python -m gown_poc.demo
make demo
python -m unittest discover -s tests
make test

The demo is stdlib-only. It builds a few mock nodes, rejects nodes that are unhealthy, incompatible, too weak, or privacy-ineligible, and selects an appropriate 12GB-friendly inference node.

Want to help?

Start with one small task:

• HANDOFF.md explains the contribution rules.
• TASKS.md lists small task cards that can be claimed from a Reddit thread.
• PATCH_PROMPT.md is a copy-paste prompt for AI-assisted patching.
• CONTRIBUTING.md explains PR expectations.

Difference from training track

training track: training/growth of experts
OWN: running and routing open-weight inference

GOWN should become the practical user-facing layer first. A future training track can later feed new experts/models into it.

Next step

Build the first real local node adapter:

llama.cpp or vLLM backend
+ NodeCard manifest
+ /v1/chat/completions-compatible gateway
+ signed health reports
+ route trace logging