● Peer-to-peer · Hyperdrive + Hyperswarm

qvac-seed

Share local AI models and LoRA adapters peer-to-peer, as pear:// links the QVAC SDK loads directly. Think BitTorrent for on-device AI: publish once, and anyone with the link downloads it from you or any other peer that has it. No Hugging Face account, no S3 bucket, no bandwidth bill.

Use it Source on GitHub

$ qvac-seed publish ./pirate-lora.gguf --meta base=QWEN3_600M_INST_Q4
pear://fd4ec0ea48…822a/pirate-lora.gguf  (lora, 1.9 MB, sha256 3b034103b34a…)
Seeding collection "default" — Ctrl+C to stop.

// on any other machine
const id = await loadModel({ modelSrc: 'pear://fd4ec0ea48…822a/model.gguf', modelType: 'llm' })

Why

Stable links

Files go into a named Hyperdrive collection, so links survive restarts and re-publishes.

No server

Seeded over Hyperswarm: Noise-encrypted connections, NAT hole-punching, no infrastructure.

Verified

Every drive carries a manifest with each file's size and SHA-256. fetch refuses a file that doesn't match.

Look before you download

ls shows what a link holds without pulling gigabytes.

In an app: Pocket Tune's Share tab

Pocket Tune publishing an adapter with qvac-seed and seeding to one peer

The receiving Pocket Tune comparing the shared adapter with the base model

Pocket Tune uses qvac-seed to share fine-tuned adapters between people: lookup in about 2 s, a 20 MB adapter downloaded and checked in 0.3 s.

Use it

npm install -g qvac-seed
qvac-seed publish model.gguf --collection my-models    # prints pear:// links, keeps seeding
qvac-seed ls pear://<key>                               # what a link holds
qvac-seed fetch pear://<key>/model.gguf -o model.gguf   # download, SHA-256 checked

Also a library (import { publishFile, fetchFile } from 'qvac-seed'). Requires Node.js 20+.