Noema Atlas
Every open model. Verified. From anywhere. The resilient way to download and preserve local LLM weights.
What Noema Atlas does
Atlas is a multi-source distribution system for open model weights. The main idea behind it is that each source provides a signed manifest for truth. The identity is the file's content hash so that same weights fetched from multiple sources all deduplicate into one content-addressed blob.
This allows Noema Atlas to treat Hugging Face as just one source among many. Atlas can intelligently use all the sources it has at its disposal for the best performance, while verifying every chunk against the file's hash as it arrives.
Verified, always
Every file is identified by its content hash (BLAKE3 + SHA-256) and described by an Ed25519-signed manifest. Corruption and tampering are caught while the bytes stream in.
Many sources, one file
The same weights from Hugging Face, an HTTPS mirror, a LAN peer, or a worldwide Iroh peer collapse into one verified blob. If a source is slow or gone, Atlas fails over automatically.
Stored once
A content-addressed cache means identical files across quants, mirrors, and sources are kept a single time. Installs are reflinks or hardlinks, never wasteful copies.
Native and lightweight
A fast, low-RAM desktop app for macOS, Windows, and Linux — no browser, no Electron. Or drive everything from the cross-platform command line.
Peer-to-peer, worldwide
Because a file's identity is its hash, Atlas can find a model anywhere. Peers announce what they hold to a shared tracker, and transfers run over Iroh — a QUIC, NAT-traversing transport that connects machines directly whenever it can. Turn on Share my models worldwide and your downloads help seed the swarm, keeping open models fast and available for everyone.
Explore the Atlas network
See the live network, learn how it works, and download Atlas for macOS, Windows, or Linux at its home on the web.
Opens in a new tab. Free and open source.