MyMedia models v3 — Chinese-CLIP text tower

Chinese-CLIP ViT-B/16 text tower — provenance and license

• Source: OFA-Sys/chinese-clip-vit-base-patch16, revision
  36e679e65c2a2fead755ae21162091293ad37834, official OFA-Sys weights.
  https://huggingface.co/OFA-Sys/chinese-clip-vit-base-patch16
  https://github.com/OFA-Sys/Chinese-CLIP
• License: MIT; see LICENSE-chinese-clip.txt with the upstream notices.
• Exported on 2026-10-06 using torch 2.14.1, transformers 5.18.0,
  ONNX 1.23.1, opset 18. fp32, single file, BERT CLS hidden state projected
  to 512 dimensions. Inputs: input_ids and attention_mask, int64 [1,52].
  Output: text_features, float32 [1,512], unnormalized.
• Same checkpoint/embedding space as the image tower in models-v2
  (image SHA256 6291342114691c97b1690645ef0a74b375fa2aa6ed60350bcf6bbf47c78f5328).
• Parity: eight zh/en queries, cosine > 0.9999 against PyTorch;
  24 text-image similarity comparisons across three deterministic synthetic
  images, maximum absolute difference 2.384185791015625e-7. See parity.json.
  App tokenizer inputs also matched the eight reference queries exactly.
• Used by MyMedia Photos for fully local, optional Labs visual search.