Qwen2-VL-72B

Qwen2-VL-72B is the largest model in Alibaba's Qwen2-VL vision-language series, with the instruct-tuned version released on September 19, 2024 alongside its quantized AWQ/GPTQ variants and the technical report. It introduced Multimodal Rotary Position Embedding (M-ROPE) to handle 1D text, 2D visual, and 3D video positional information, and could analyze videos over 20 minutes long — a standout capability at release. The 2B and 7B siblings shipped under Apache 2.0; the 72B ships under Alibaba's Qwen license.

The honest picture: the Qwen license is not OSI-approved open source, so commercial users should read the terms rather than assume Apache-style freedoms. It is a 2024 model, superseded by Qwen2.5-VL and Qwen3-VL on most benchmarks. But the weights are downloadable, it integrates with transformers and vLLM, and at release it was among the strongest open multimodal models available.