ColorBench: leaderboard

Multiple-choice questions (3 to 6 options) about colour in images: recognising, extracting, comparing and counting colours, colour illusions, camouflage and colour-blindness plates; 955 questions per model, accuracy in percent. Unlike general image QA boards, every item turns on colour.

Metric: P&R Overall accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 31 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.0 Flash55.4#723
2llava-onevision-qwen2-72B-ov-sft51.9#1957
3InternVL2.5-38B50#1366
4InternVL2.5-78B48.8#1036
5InternVL2.5-26B46.8#2191
6LLaVA-v1.6-34B46.6#5432
7InternVL2-26B46.3#2639
8Qwen 2.5 VL 7B Instruct46.2#1893
9InternVL2-40B45.6#2094
10Cambrian-34B45.3#5007
11InternVL2.5-8B45.2#2401
12LLaVA-NeXT-72B45.2#3206
13llava-onevision-qwen2-7B-ov44.7#3252
14InternVL2-Llama3-76B44.6#2021
15eagle-x4-8B-plus44.1

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=colorbench · How It Works · Data refreshed daily, snapshot 2026-11-17.