Atlas

Models

← All models

Qwen2 VL 7B Instruct

Alibaba · 2024-08-29 · 8.3B parameters

Qwen2-VL 7B Instruct is Alibaba's roughly 8.3-billion-parameter instruction-tuned vision-language model (about a 7.6B language model plus a 675M vision encoder) from the Qwen2-VL series, released in 2024. It combines a Qwen2 language backbone with a visual encoder supporting dynamic resolutions and long-context processing, enabling document analysis, OCR, chart understanding, temporal reasoning over video, and structured output generation.

Benchmark scores

BenchmarkScore
Arena Vision — Chinese (No Style Control)999
Arena Vision — Chinese (Style Controlled)1038
Arena Vision — English (No Style Control)1006
Arena Vision — English (Style Controlled)1040
BALROG3.7
BALROG — BabaIsAI (LLM)7.6
BALROG — BabaIsAI (VLM)11.9
BALROG — BabyAI (LLM)4.0
BALROG — BabyAI (VLM)2.0
BALROG — Crafter (LLM)6.4
BALROG — Crafter (VLM)5.5
BALROG — MiniHack (LLM)2.5
BALROG — MiniHack (VLM)5.0
BALROG — NetHack (LLM)0.0
BALROG — NetHack (VLM)0.0
BALROG — TextWorld (LLM)1.6
BBH (Open LLM Leaderboard v2)54.6
Capability109.0
CharXiv Reasoning34.6
CharXiv-D58.0
GPQA (Open LLM Leaderboard v2)32.0
IFEval (Open LLM Leaderboard v2)46.0
MATH Level 5 (Open LLM Leaderboard v2)19.9
MATH-Vision16.3
MMLU-Pro (Open LLM Leaderboard v2)40.9
MuSR (Open LLM Leaderboard v2)43.8
Vision Arena Overall1031
Vision Arena Overall No Style Control990
Loading Atlas data…