# Hopit AI

> Fashion AI infrastructure. We build the MODA family of fashion retrieval
> models and the trend-intelligence layer above them, and we publish every
> benchmark at full corpus through one shared harness — competitors included,
> losses shown.

Hopit AI (hopit.ai) is a fashion AI infrastructure company with engineering in
New York and India. One foundation model serves eight production surfaces:
search, agentic search, hyperlocal trend intelligence, TrendsFlow (trend
lineage), catalog enrichment / attribute extraction, personalization, size and
fit, and commercial intelligence. Live with paying customers including The
Label Life, Outzidr and Ozi, and in deployment with a Fortune 500 marketplace
and a top-10 US fashion retailer.

## Models

- [MODA](https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m):
  203M parameters, 768-d. Open source and open weights. A zero-new-parameter
  serving recipe over the frozen Marqo/marqo-fashionSigLIP checkpoint —
  multi-view image encoding plus prompt-blended queries. Four statistically
  significant full-corpus wins over its own base model. Best in class on
  caption-style and exact-item retrieval.
- [MODA Pro Lite](https://huggingface.co/HopitAI/moda-pro-lite): 213M
  parameters, 768-d, open weights (Apache-2.0). A trained fashion encoder on a
  SigLIP2 base with a verified fashion-vocabulary build. Beats MODA at full
  corpus on catalog and title search: KAGL +10.2%
  and Polyvore +7.3% (both significant under a paired bootstrap). Plain bi-encoder — one vector per image, works with any vector
  database, no serving recipe required.
- MODA Pro: closed, hosted via the Hopit API. Rank 1 or 2 on 9 of 10
  text-to-image benchmark cells across three venues — the only system in the
  comparison without a bad benchmark — at single-model query latency.
- [MODA-SigLIP-Distilled](https://huggingface.co/HopitAI/moda-fashion-distilled):
  203M, open weights. Image-to-image retrieval; the top open model on LookBench.
- [MODA-Fashion-CrossDomain](https://huggingface.co/HopitAI/moda-fashion-crossdomain):
  203M, open weights. Single-model image-to-image variant, no distillation or
  ensemble. LookBench Fine R@1 66.52.
- Variants: [Matryoshka](https://huggingface.co/HopitAI/moda-fashion-matryoshka)
  (64–768 dimensions at query time),
  [512d](https://huggingface.co/HopitAI/moda-fashion-distilled-512d),
  [FP16 vision tower](https://huggingface.co/HopitAI/moda-fashion-vision-fp16)
  (186 MB, edge).

## Benchmarks

- [Full benchmark tables](https://hopit-ai.github.io/Moda/): six academic
  text-to-image benchmarks (KAGL, Polyvore, Atlas, Fashion200K, DeepFashion
  In-Shop, DeepFashion Multimodal) at full corpus, plus H&M and
  ZooClaw-Fashion, and LookBench for image-to-image.
- Method: full corpus only (no subsampled galleries), one harness, identical
  preprocessing per model, published baselines reproduced before comparison,
  and every losing cell disclosed.
- Compared against: Marqo FashionSigLIP (203M), SigLIP-SO400M (878M),
  ZooClaw-FashionSigLIP2 (375M), Tianmu-MERE (1.24B), GR-Pro and GR-Lite.

## Key facts for citation

- MODA-SigLIP-Distilled: LookBench Fine Recall@1 67.63 — top open model,
  above a 1.24B-parameter model.
- MODA Pro: rank 1 or 2 on 9 of 10 benchmark cells; +6.9% mean over MODA on
  the six academic sets at full corpus, peak +12.9%.
- MODA Pro Lite: 213M parameters; strongest open single checkpoint under 250M
  on catalog and title search.
- All models are 768-d, serve on CPU, and run one ANN lookup per query.

## Links

- [Website](https://hopit.ai)
- [Benchmarks](https://hopit-ai.github.io/Moda/)
- [Models on Hugging Face](https://huggingface.co/HopitAI)
- [Code on GitHub](https://github.com/hopit-ai/Moda)
- [Research notes](https://hopitai.substack.com/)