Perception Encoder
MetaMeta FAIR's family of contrastively trained vision encoders. The PE Core models are CLIP-style image and video encoders with a matching text encoder, for zero-shot classification and retrieval, in sizes from T/16 to G/14, the smaller ones distilled from the largest. Meta describes them as outperforming SigLIP 2 on image benchmarks and InternVideo2 on video.
Only the PE Core checkpoints are declared. PE Lang and PE Spatial, tuned for language models and dense prediction, and the audio-visual PE-AV models come from the same repository but are not retrieval models; the Perception LM checkpoints are a separate entry.
Openness
3 high confidence- weights
- open (five PE Core checkpoints on the Hub, ungated)
- data
- closed (the image-text pretraining corpus is not named or released; the PE Video Dataset used for video fine-tuning is released)
- code
- partial (model code with zero-shot classification and retrieval evaluation; no pretraining pipeline)
- license
- Apache-2.0 (OSI)
The PE Core weights download freely under Apache-2.0. Meta releases the video dataset used to fine-tune them, but not the image-text corpus they were pretrained on or the pipeline that trained them.
- https://arxiv.org/abs/2504.13181 recorded 2026-11-12
Abstract: "To foster further research, we release our models, code, and novel dataset of synthetically and human-annotated videos"; the image pretraining corpus is not named.
- https://huggingface.co/api/models/facebook/PE-Core-G14-448 recorded 2026-11-12
Hub record: "gated":false, cardData "license":"apache-2.0", PE-Core-G14-448.pt in the file list.
- https://huggingface.co/api/models/facebook/PE-Core-L14-336 recorded 2026-11-12
Hub record: "gated":false, cardData "license":"apache-2.0".
- https://raw.githubusercontent.com/facebookresearch/perception_models/main/apps/pe/README.md recorded 2026-11-12
PE README: "All checkpoints released on this page, unless otherwise specified, are released with the [Apache 2.0 license]"; "PE core is our base model trained with our robust image pretraining schedule and finetuned on the data generated by our synthetic video data engine"; the PE Video Dataset is released; usage covers feature extraction and zero-shot evaluation.
- https://raw.githubusercontent.com/facebookresearch/perception_models/main/README.md recorded 2026-11-12
Repository README: badges "Code License: Apache 2.0" and "Model License: Apache 2.0"; "[Jul-11-25]: We have release 8 new checkpoints for Perception Encoder: 2x small core models (T and S)".
Adoption
3 high confidenceHugging Face downloads over the trailing month, summed across the five PE Core checkpoints and led by the L/14 model. The timm conversions are excluded as format copies.
- https://huggingface.co/api/models?author=facebook&search=PE-&limit=100 recorded 2026-11-12
"downloads": PE-Core-L14-336 259027, PE-Core-B16-224 26391, PE-Core-G14-448 21253, PE-Core-S16-384 1326, PE-Core-T16-384 876; the *_timm repos draw 0.
Capability
4 medium confidenceMeta's figures put PE Core slightly ahead of SigLIP 2 on the same zero-shot ImageNet and COCO retrieval tests. The margin is narrow and comes from Meta's own table, so it sits level with Google's encoder rather than above it.
- https://raw.githubusercontent.com/facebookresearch/perception_models/main/apps/pe/README.md recorded 2026-11-12
PE core table, IN-1k / COCO-T2I: "**G/14** 448px | PE-Core-G14-448 | 85.4 | ... | 58.1"; "**L/14** 336px | PE-Core-L14-336 | 83.5 | ... | 57.1"; "**B/16** 224px | ... | 78.4 | ... | 50.9"; "PE core can outperform SigLIP2 on Image CLIP and InternVideo2 on Video CLIP".
Verified 2026-11-12