Datasets Models Results
Benchmark Report 2026

Visual Search &
Parts Identification Benchmark

A comprehensive evaluation of state-of-the-art embedding models for visual product search. Comparing 11 models across 6 diverse datasets spanning automotive parts, hardware, and industrial components.

View Results Read Report
6
Datasets
11
Models
60+
Evaluations
3
Metrics

Key Findings

Highlights from our benchmark evaluation

Top Performer
GEM v5.1 (ours)

Achieves highest average R@1 and mAP@20 across all datasets
Best on Stanford Online Products
94.17%

Recall@5 achieved by GEM v5.1 (ours)
Hardest Dataset
Clips-and-Connectors v1

Industrial parts with highest variance in model performance

Explore the Benchmark

Navigate through different aspects of our evaluation study.

Datasets

Explore the 6 diverse datasets used in our benchmark, including industry and academic benchmarks.

Automotive v1 Stanford Online Products ILIAS +3 more

Models

Compare 11 state-of-the-art embedding models from 6 providers including Google, Meta, Cohere, and Jina AI.

nyris Google Meta +3 more

Results & Analysis

Interactive charts, comparison tables, and detailed analysis of model performance across all datasets.

Charts Tables Rankings