shardr¶
shardr is a decentralized LLM repository with sync-based distribution. It keeps large language models available and digitally sovereign: model files live in content-addressed storage on your own machines, synchronize over a BitTorrent-based peer network, and are served through an OpenAI-compatible runtime. Availability grows with every node that participates; no single provider is required.
The system has two parts. shardhive is the storage daemon: a content-addressed store (CAS) with imports from local files, Hugging Face, and BitTorrent, plus the sync client and API v1 over a mode-0600 Unix socket. shardr is the model runtime and management CLI: it runs a model with layered runtime configuration and maps the weights directly out of the CAS at serve time.
Content is addressed by digest and verified on every write, independent of
how a byte arrived. The reference ns/name:quant addresses an artifact;
the quant vocabulary is protocol-level and parses identically on every
node.
Where to go next¶
| Goal | Path |
|---|---|
| Import and manage models | Importing models |
| How seeding and peer synchronization work | Swarm & seeding |
| Configure shardhive and the model runner | Configuration |
| Work on the code (architecture, layout, conventions) | Developer Guide |
| Concept tour and quickstart (in progress, Epic #41 S5) | Get Started |
Status¶
Operational end to end today: import → CAS → synchronization → serving a
real GGUF with the pinned llama.cpp runtime. Specifications live in
docs/specs/ in
this repository and are canonical; site pages summarize them and link to
the code tree rather than duplicating them.