LocalFold browser weights

The model parameters localfold.org loads in the browser. Everything here is an export, packed for WebGPU: tensors are quantised, laid out in the dense atom layout the shaders read, and split into shards a page can fetch in parallel. They are not checkpoints for training and have no optimiser state.

Source: https://github.com/sokrypton/localfold

What is here

directory	model	encoding	size
af2-monomer/	AlphaFold 2, model 1 PTM	int8, float16 block scales	227 MB
af3-int5/	AlphaFold 3 diffuser: trunk, diffusion head, confidence head	int5, asymmetric per-block	265 MB

Each directory carries a manifest.json naming every tensor, its dtype, its shard and its byte offset. The page compiles that manifest in and fetches only the shards.

Licences differ per directory

🔴 af3-int5/ holds DeepMind's AlphaFold 3 parameters, which carry the AlphaFold 3 Model Parameters Terms of Use and its Prohibited Use Policy. Downloading them means accepting those terms. They are not OpenFold3's, whatever the folder name suggests.

af2-monomer/ is an export of AlphaFold 2's published parameters, which are released under CC BY 4.0.

If you are unsure which terms apply to what you intend to do, read the upstream terms rather than this table.

Using them

The page resolves a shard against a pinned revision, so a URL is immutable:

https://huggingface.co/sokrypton/localfold/resolve/<commit-sha>/af3-int5/

Pinning a commit rather than main is deliberate: a shard that changes under a manifest that did not is a fold that reads tensors at the wrong offsets and returns noise.