SEA-CLIP-Tiny

A 46.11M-parameter multilingual text-vision embedding model for English and seven Southeast Asian languages (Indonesian, Javanese, Sundanese, Malay, Thai, Vietnamese, Burmese), distilled from MetaCLIP2-ViT-B-16-worldwide.

Across the seven SEA languages it reaches 12.9 / 31.5 / 42.2 average R@1 / R@5 / R@10 — the strongest average among the compact students evaluated in the paper — improving average R@10 over MobileCLIP2 by 12.1 points with 38.4% fewer parameters and lower CPU latency.

Usage

import open_clip

model, _, preprocess = open_clip.create_model_and_transforms('hf-hub:fassabilf/sea-clip-tiny')
tokenizer = open_clip.get_tokenizer('hf-hub:fassabilf/sea-clip-tiny')

Model

Architecture	ViT-T/16 vision tower + 12-layer / 384-wide text tower, embed dim 512
Tokenizer	CLIP BPE, vocab 49408, context length 77
Parameters	46.11M (5.62M vision + 40.49M text)
Training data	CC12M + CulturalGround-OE-filt + WIT + Bloom + Mammoth-VL-SEA (12.72M pairs)
Teacher	MetaCLIP2-ViT-B-16-worldwide

Results

Retrieval R@1 on the held-out splits of each training source, zero-shot ImageNet accuracy, and the paper's retrieval-only Avg@1 over XM3600, Flickr30k-200 and XTD-200 (%).

CG R@1	WIT R@1	Bloom R@1	ImageNet	R@1-Avg
29.3	21.8	23.3	36.4	13.9

Reproduce

Training and evaluation code: https://github.com/fassabilf/sea-clip-tiny. The exact training configuration of this checkpoint is in params.txt in this repo.

Citation

@inproceedings{seacliptiny2026,
  title     = {SEA-CLIP-Tiny: Efficient Multilingual Text-Vision Embedding for Southeast Asian Languages},
  booktitle = {Asian Conference on Computer Vision (ACCV)},
  year      = {2026}
}