---
license: mit
tags:
- mirror
- xai
- multimodal
- clip
- zero-shot-image-classification
---

# CLIP ViT-B/32 (OpenAI) — xaitalk mirror

This is a **bit-identical mirror** of the canonical artifact from OpenAI.

The mirror exists only as a resilience fallback for the
[xaitalk](https://github.com/xaitalk/xaitalk) library — the upstream remains
authoritative. All credit and licensing for the model belong to the original
authors.

## Attribution

| Field | Value |
|---|---|
| **Original authors** | OpenAI |
| **Upstream (authoritative)** | https://huggingface.co/openai/clip-vit-base-patch32 |
| **Source repo** | https://github.com/openai/CLIP |
| **Paper** | https://arxiv.org/abs/2103.00020 (Radford et al. 2021) |
| **License** | `mit` (inherited from upstream — please respect upstream's terms) |
| **Mirror file** | `pytorch_model.bin` |
| **SHA-256** | `a63082132ba4f97a80bea76823f544493bffa8082296d62d71581a4feff1576f` |
| **Size** | 605,247,071 bytes (577.2 MB) |

## How xaitalk loads this file

```python
from xaitalk.hub import ensure_model
weights_path = ensure_model("clip-vit-b32")
# Tries the canonical upstream first; falls back to this xaitalk mirror
# automatically if upstream is unreachable.
```

## Why mirror?

xaitalk's research-grade reproducibility claim relies on every weight file
being recoverable years from now. We mirror artifacts ≤ 2.5 GB under
`xaitalk/*-mirror` so the pipeline survives upstream URL changes, repo
renames, or deletions. Bit-level parity with the canonical is asserted in
CI via `python -m xaitalk.hub verify-mirrors`.

## Citation

If you use this model, **please cite the original paper** (not the mirror):

```
https://arxiv.org/abs/2103.00020 (Radford et al. 2021)
```