Qwen2-VL-72B Instruct

À propos de Alibaba DAMO Academy (Qwen Team)

Fondée 2017 · Hangzhou, China

The Qwen (Tongyi Qianwen) team sits inside Alibaba Cloud's DAMO Academy, the company's research arm founded in 2017 in Hangzhou. The team is led by Junyang Lin and Le Hou and counts dozens of researchers across NLP, vision and speech. Qwen has produced one of the most prolific open-source model lines in the world, including Qwen-1.5, Qwen2 (June 2024), Qwen2.5 (September 2024), the Code, Math, Audio and VL (vision-language) families, and the December 2024 release of Qwen2.5-VL. Qwen2-VL launched in August 2024 in 2B, 7B and 72B sizes, all released on Hugging Face and ModelScope; the 72B Instruct variant became one of the top open-weights vision-language models worldwide, frequently matching closed-source peers on OCR-heavy benchmarks like DocVQA and ChartQA. Alibaba offers Qwen models commercially through Alibaba Cloud and Bailian.

Architecture

Decoder-only Transformer with Naive Dynamic Resolution Vision Transformer

Qwen2-VL-72B-Instruct combines the Qwen2 72B decoder-only Transformer with a custom 675M ViT vision encoder using Naive Dynamic Resolution: instead of resizing every image to a fixed grid, the encoder accepts the native resolution and generates a variable number of visual tokens per image. The model also introduces Multimodal Rotary Position Embedding (M-RoPE) that encodes positions in time (for video), height and width separately, enabling single-stream multimodal video understanding. The model supports up to 20 minutes of video input via uniform frame sampling, single-frame image input at variable resolution up to ~16K visual tokens, and a 131,072-token text context window. Training proceeded in three stages: contrastive vision-language pretraining, multimodal pretraining on interleaved image-text and video-text data, and supervised fine-tuning with chain-of-thought multimodal instructions. Weights are released under the Qwen licence (free for commercial use under specific terms).

Paramètres

72B (~73B with vision encoder)

Contexte

131 072 tokens

Capacités

• Open-weights 72B vision-language model under permissive Qwen licence
• Naive Dynamic Resolution: native image aspect ratio without fixed grid
• Multimodal Rotary Position Embedding (M-RoPE) for joint image and video
• Up to 20 minutes of video understanding
• 131K-token text context
• Top open-weights scores on DocVQA, ChartQA, MathVista, RealWorldQA
• Strong OCR across English, Chinese, Japanese, Korean and European languages
• Best for: open-weights document AI, video QA, OCR-heavy multilingual workloads

Entraînement et licence

Multi-stage curriculum: contrastive vision-language pretraining on large web image-text pairs, multimodal pretraining on interleaved image-text and video-text data, supervised fine-tuning on curated chain-of-thought multimodal instructions.

Licence: Qwen Licence (commercial use permitted under 100M MAU; bespoke licence required above).

Tests de sécurité: Alibaba publishes a model card with safety evaluations and integrates Tongyi safety filters in cloud deployments; no separate full red-team report.