README.md

---
license: apache-2.0
library_name: jax
tags:
  - protein-structure-prediction
  - alphafold3
  - jax
  - equinox
  - biology
---

# Protenij — JAX/Equinox weights for Protenix

This repository hosts JAX/Equinox-converted model weights (and a mirror of the
original PyTorch `protenix-v2` checkpoint) for use with
[protenij](https://github.com/escalante-bio/protenij), a JAX/Equinox translation
of [Protenix](https://github.com/bytedance/Protenix), ByteDance's implementation
of the AlphaFold 3 architecture.

The JAX/Equinox weights are format conversions of the original PyTorch
checkpoints released by ByteDance — the underlying model parameters are
numerically identical, only the serialization format has changed (PyTorch `.pt`
→ Equinox `.eqx` + pickled skeleton).

## Files

| File | Format | Size | Source |
| --- | --- | --- | --- |
| `protenix-v2.eqx` / `protenix-v2.skeleton.pkl` | Equinox | 1.86 GB | Converted from `protenix-v2.pt` |
| `protenix-v2.pt` | PyTorch | 1.86 GB | Mirror of upstream ByteDance release |
| `protenix_base_default_v1.0.0.eqx` / `.skeleton.pkl` | Equinox | — | Converted from upstream |
| `protenix_base_20250630_v1.0.0.eqx` / `.skeleton.pkl` | Equinox | — | Converted from upstream |
| `protenix_mini_default_v0.5.0.eqx` / `.skeleton.pkl` | Equinox | — | Converted from upstream |
| `protenix_tiny_default_v0.5.0.eqx` / `.skeleton.pkl` | Equinox | — | Converted from upstream |
| `components.v20240608.cif` | Data | — | CCD chemical components (upstream) |
| `components.v20240608.cif.rdkit_mol.pkl` | Data | — | CCD rdkit mol cache (upstream) |
| `clusters-by-entity-40.txt` | Data | — | PDB entity-40 clusters (upstream) |

## Usage

```python
from protenix.backend import load_model
model = load_model("protenix-v2")  # auto-downloads from this repo
```

## License and attribution

Released under the **Apache License 2.0**, matching the upstream
[bytedance/Protenix](https://github.com/bytedance/Protenix) project.

The upstream Protenix README explicitly states:

> "The Protenix project including both code and model parameters is released
> under the Apache 2.0 License. It is free for both academic research and
> commercial use."

### Modification notice (Apache 2.0 §4(b))

The `.eqx` and `.skeleton.pkl` files in this repository are **format
conversions** of the original PyTorch checkpoints released by ByteDance. The
PyTorch state dicts were loaded and the tensors re-serialized in Equinox format
using
[`protenix/backend.py`](https://github.com/escalante-bio/protenij/blob/main/protenix/backend.py)
and
[`translate_models.py`](https://github.com/escalante-bio/protenij/blob/main/translate_models.py).
No weights were retrained, fine-tuned, or otherwise numerically modified.

The `protenix-v2.pt` file in this repository is a bit-for-bit mirror of the
original PyTorch checkpoint hosted at
`https://protenix.tos-cn-beijing.volces.com/checkpoint/protenix-v2.pt` (mirrored
here after the upstream URL became unreachable).

### Copyright notice (Apache 2.0 §4(c))

Copyright 2024 ByteDance and/or its affiliates. The original Protenix code and
model parameters were released under Apache License 2.0. See the `LICENSE` file
in this repository for the full license text.

### Citations

If you use these weights, please cite the original Protenix work:

- Protenix repository: https://github.com/bytedance/Protenix
- Protenix technical reports in `docs/` of the upstream repository

## Disclaimer

These files are provided as-is. The weights are format conversions only — for
the authoritative source and for training code, model cards, and technical
reports, refer to the [upstream ByteDance Protenix
repository](https://github.com/bytedance/Protenix).