The intelligence layer fashion commerce runs on.

hopit.ai · Benchmarks · Models · Research

We build fashion-native retrieval and trend intelligence — and we measure it in public. Every number below is full corpus, one harness, competitors included, losses shown.

Where they stand. MODA-SigLIP-Distilled is the top open model on LookBench image retrieval, above a 1.24B-parameter model. On text-to-image, MODA Pro Lite+ leads every model at ≤250M parameters on catalogue search (KAGL +10.9%, Polyvore +8.7% over MODA), and MODA Duo routes each query to whichever open model suits its shape, beating both on mixed traffic. Our hosted MODA Pro is rank 1 or 2 on five of six full-corpus benchmarks — an 878M model beats us on four of them, and that is in the tables too.

Every cell is MAP@10 at full corpus under one evaluator (pytrec_eval map_cut.10), competitors included.

Full tables, including every cell we lose: hopit-ai.github.io/Moda

• Moda — the open benchmark and model family: harness, evaluation code, and the write-ups, including the experiments that failed.
• india-trade-cli — agentic research over Indian equities. Different domain, same conviction: publish the method, measure the result.

• Full corpus only. No subsampled galleries; screening runs are never mixed with full-corpus rows.
• One harness. Ours and competitors' models run identical preprocessing and protocol — we reproduce published baselines before comparing against them.
• Losses shown. Every model card links the benchmarks it loses.
• Negative results published. Three training approaches failed before the fourth worked; all three are written up.

_{Made with ♥ in NYC and India}