LeVJEPA 1.5 Large
What it is
LeVJEPA 1.5 Large is a self-supervised video encoder from the galilai-group model card. The model card reports 303 million parameters and positions it for frozen feature extraction, retrieval, attentive or linear probing, and downstream heads. It does not include a classification head.
Reported results
The model card reports a mean score of 78.76 across 22 image and video datasets, using about 5,300 unique video hours and 151 exaflops. Its comparison lists V-JEPA 1 at 79.77 with about 136,000 video hours and 844 exaflops, and V-JEPA 2 at 81.63 with an estimated 1.7 million hours and about 4,300 exaflops. These are model-card results, not an independent reproduction.
Method and use
The model uses a single encoder with an invariance loss over global and local clip views, regularized with SIGReg. The model card says it does not use an EMA target encoder, predictor, stop-gradient, or pixel decoder. Its evaluation uses frozen encoders with attentive probes.
The documented input is 16 ImageNet-normalized frames at 224 by 224 pixels, sampled at about 7.5 frames per second. The Hugging Face implementation requires trust_remote_code=True; the card also documents a legacy encoder path.
Limitations and checks
- The model card reports a lower aggregate score than V-JEPA 1 and V-JEPA 2; the comparison is about data and compute efficiency, not a higher benchmark score.
- The listed license is CC BY-NC 4.0, which is non-commercial.
- Review the shipped modeling code before enabling
trust_remote_code=True. - The evaluation figures and compute estimates have not been independently reproduced here.
Related
Pending Jean review
- ai-ecosystem/levjepa-1-5-large —
ai_draft; keep as a candidate until the model-card and paper results are independently reviewed.