Open weights shouldn't have an owner.
Today one company hosts most of them. Morcom replaces the host with a signed index and a swarm: labs, universities and anyone else seed the bytes, and Morcom only says which bytes are the real ones.
$ export HF_ENDPOINT=https://hub.morcom.ai- anchor-eth-10
- anchor-utoronto0
- seed-070
- mirror-https0
Three steps. The third only runs when the swarm is empty.
- Resolve
- Your client asks hub.morcom.ai for a repo and gets a signed manifest back: SHA-256 for every file, the signer's identity, the licence. Manifests are OpenSSF Model Signing bundles on Sigstore, so existing verifiers already work.
- Fetch
- Blocks arrive over BitTorrent v2 from whoever is seeding. Each 16 KiB block is checked against the manifest's Merkle root as it lands. A bad peer costs you bandwidth, not integrity.
- Fall back
- Every repo has at least one signed HTTP mirror. If no peers answer, or your firewall blocks them, the client gets a plain range-request download. Never slower than today.
Three services. None of them store weights.
Everything Morcom operates is small, cheap and self-hostable. The heavy bytes never pass through us.
- Index
- Postgres and Meilisearch. Which repos exist, their files and hashes, who signed them, who is seeding right now.
- Tracker
- Introduces clients to peers. Metadata only; a few hundred dollars a month of bandwidth at scale.
- Signing service
- Verifies manifests and appends every registration to a public transparency log.
- Weight store
- Not operated. A repo on Morcom is a signed manifest and a list of places to get the bytes.
Kept online by the institutions that use it
The Debian mirror model. Academic Torrents has moved terabytes a day this way since 2013. Seeding is public, measured and credited.
- Anchors
- Universities and labs that commit to seed a defined model set under an availability agreement. Anchors hold seats in governance.
- Seed operators
- Anyone running the seed daemon. Uptime, bytes served and pinned models appear on your profile and on every model page you keep alive.
- Fallback store
- Repos with zero live seeders are served from a foundation-run bucket on zero-egress storage. It is the only bytes Morcom pays for.
- Seed daemon
- A packaged libtorrent service. Configure a pin set, point it at storage, and it reports its own metrics to the index.
Malware, licences and takedowns
Scanning is necessary and not sufficient. Knowing who signed a model is the defence Morcom is built around.
- Safetensors preferred
- Pickle files are flagged and quarantined. Ingestion runs picklescan and ClamAV; neither catches everything, and we say so.
- Licence required
- Every manifest carries an SPDX licence field. No field, no listing.
- Delist, not delete
- Morcom cannot remove bytes from a swarm. It can remove a manifest from the index, through a public notice-and-delist process.
- Names are retired
- A deleted organisation name is never reissued, so nobody can ship a lookalike under a trusted name.
Egress is the cost. Morcom doesn't pay it.
300 TB of fallback storage, per month, against the same set served once from a conventional bucket.
- Backblaze B2
- Storage only; egress free via Cloudflare
- ≈ $1,800
- Cloudflare R2
- Storage only; zero egress
- ≈ $4,500
- S3 + CloudFront
- Egress alone, one full read of the set
- $15,000 – 27,000
Order-of-magnitude figures from published pricing. Index, tracker and signing add low thousands, and shrink as mirrors take them over.
If our servers vanish, the index does not.
Being copyable is a requirement. Any mirror can rebuild the full signed index without asking us.
- Append-only log
- Every manifest registration is recorded in a signed transparency log. Nobody rewrites history, including Morcom.
- Signed snapshots
- Published on a schedule, replicated git-style and over torrent. The index travels the way the models do.
- Multi-party root keys
- TUF roles over the snapshots; root keys held by distinct member organisations, as PyPI does. A compromised mirror cannot serve stale or forged state.
Seed the first set.
Morcom is being built in the open by Orchestrator AI, a forward-deployed AI engineering firm in Calgary. Foundation handoff is planned once the first anchors are seeding, and the first anchors write the charter.