Atlas

Benchmarks

← All benchmarks

CV-Bench

Multimodal · 2024-06-24

CV-Bench, from Cambrian-1, repurposes ADE20K, COCO and Omni3D annotations into 2,638 vision-centric multiple-choice questions: 2D counting and spatial relation (1,438) and 3D depth order and relative distance (1,200). The official score averages 2D accuracy (itself the mean of the ADE20K and COCO accuracies) with 3D accuracy rather than pooling items.

Top models (higher is better)

ModelScore
LFM3-VL-2B88.8
Qwen3.5-4B88.4
Gemma 4 12B IT87.5
Qwen3.5-2B84.3
Gemini 3.5 Flash-Lite82.7
Gemma 4 E2B IT77.6
GPT-5.6 Luna76.9
Loading Atlas data…