About us
Four hours to move a checkpoint, every single time
Weightsmith was built by a team that spent more time moving weights than training them.
Origin
The file transfer was the bottleneck
Every experiment ended with a 140 GB upload and every deployment started with a 140 GB download, over tooling designed for source code. Content-addressed chunking fixed both, and once artefacts were addressable it became obvious that provenance should be attached to them rather than written in a README.
- Content-addressed, so identical tensors are stored once
- safetensors and GGUF only — no code execution on load
- Public models are free to host and free to pull
An internal chunk store
One team, one very slow link.
Public registry
Open models hosted free.
Signed provenance
Training manifests attached to artefacts.
Torrent seeding
Release-day pulls stop melting the origin.
128,400 versions
4.2 PB physical, 11.8 PB logical.
How we work
Boring on purpose
We would rather run software that is three years old and understood than something new that pages us at 04:00.
Own the path
We buy transit from multiple carriers and peer directly wherever it is possible, so a single bad hop is never a single point of failure.
Write it down
Every incident gets a public write-up. Every design decision gets an RFC. Institutional memory beats heroics.
Some of the people behind it
A quarter of the company has been here since the first rack.
We are hiring
Remote-first across EU time zones, with a small office network for people who prefer one.