—
files shared
Download and share local LLM weights the resilient way — content-addressed, cryptographically signed, and peer-to-peer worldwide. Hugging Face is one source, not the only one.
The local-LLM community keeps asking for the same things: don't depend on one host, publish hashes, make it fast, and reuse what's already on disk. Atlas does exactly that.
Every artifact is identified by its content hash (BLAKE3 + SHA-256) and described by an Ed25519-signed manifest. Corruption and poisoning are caught while bytes stream in; bad sources are banned.
The same weights from Hugging Face, an HTTPS mirror, IPFS, or a worldwide Iroh peer all collapse into one verified blob. If one source is slow or gone, Atlas fails over automatically.
A content-addressed cache means identical files across quants, mirrors, and sources are stored a single time; installs are reflinks/hardlinks, not copies.
If it's local and verified, access can't be silently altered or withdrawn. Seed what you download and help preserve open models for everyone.
Type a name and browse live results from Hugging Face — downloads, likes, license, quant variants.
Choose a quant. Atlas shows every place it can come from, and its Content ID — the file's P2P address.
Pulled from Hugging Face or a nearby/worldwide peer, checked against its hash byte-for-byte, and deduplicated.
Because identity is the content hash, Atlas can find a model anywhere. Peers announce what they have to this tracker; transfers run over Iroh (QUIC, NAT-traversing) — directly between machines when possible. Turn on Share my models worldwide and your openly-licensed downloads help seed the swarm. Gated/licensed and privately-imported models stay on your machine unless you opt them in — Atlas verifies content, not licenses, so you decide (and are responsible for) what you redistribute.
content-id (sha256): a few bytes → the same file, everywhere · verified, whoever serves it
Both apps are built on the same verified, peer to peer core and read from the same library, so your models, your shares, and your settings stay in step no matter which one you open. What sets them apart is the interface and what it asks of your machine.
The lean desktop app, drawn natively with no browser and no Electron underneath it. It stays light on memory while still doing everything the project offers, which makes it the comfortable choice for almost everyone.
Building from source or just curious? Everything lives on GitHub.
Sign a manifest with your Ed25519 key and publish it here. Anyone can then fetch your model from any source and prove it's exactly what you released. Atlas is open source and built on a shared Rust core that also powers the desktop, CLI, and mobile apps.