Atlas

Benchmarks

← All benchmarks

MMLongBench-Doc

Multimodal · 2024-07-01

MMLongBench-Doc evaluates long-context document understanding with 1,082 expert-annotated questions over 135 lengthy PDF documents (about 47 pages on average) whose evidence may span text, tables, charts, images and layout, including unanswerable questions. Scores report overall answer accuracy under the publisher's own page-rendering setup. Higher is better.

Top models (higher is better)

ModelScore
Qwen3.8-Flash-Next67.9
GLM-5.3 Flash65.8
Nemotron 3.5 Super VL 120B A12B BF1662.8
Nemotron 3.5 Super VL 120B A12B NVFP462.8
Qwen3.6 Plus (2026-04-02)62.0
Opus 4.561.9
Qwen3.5 397B A17B61.5
Gemini 3 Pro Preview60.5
Qwen3.5-27B60.2
Qwen3.5 Plus (2026-02-15)59.7
Qwen3.5 35B A3B59.5
Kimi K2.558.5
Qwen3.5-9B57.7
Nemotron 3 Nano Omni 30B A3B57.5
Qwen3.5-Omni-Plus57.5
Qwen3-VL-235B-A22B-Instruct57.0
Qwen3 VL 235B A22B Thinking56.2
Gemini 2.5 Pro55.6
Qwen3 VL 32B Instruct55.4
Sonnet 455.4
Qwen3 VL 32B Thinking54.6
Opus 4.154.5
Qwen3.5-4B54.2
Qwen3.5-Omni-Flash53.6
GPT-551.5
GPT-5 Mini50.3
Inkling-Small49.1
Gemini 2.5 Flash49.0
Qwen3 VL 8B Thinking48.0
Qwen3-VL-8B-Instruct47.9
Qwen3 VL 30B A3B Thinking47.4
Qwen3 VL 30B A3B Instruct47.1
Gemini 2.5 Flash-Lite46.5
Qwen3.5-2B45.4
GLM-4.5V44.7
Qwen3 VL 4B Thinking44.4
Qwen3 VL 4B Instruct43.5
Kimi-VL-A3B-Thinking-250642.1
Kimi-VL-A3B-Instruct35.1
Qwen3 VL 2B Thinking33.8
Loading Atlas data…