Alibaba Cloud’s Tongyi Qianwen opens sources of new visual models, Qwen2.5-VL and Qwen2.5-1M. Qwen2.5-VL comes in three sizes: 3B, 7B, and 72B, and the flagship version Qwen2.5-VL-72B has won the visual comprehension championship in 13 authoritative tests, surpassing GPT-4o and Claude3.5.

The new Qwen2.5-VL can more accurately analyze image content, providing breakthrough support for more than an hour of video comprehension. Without fine-tuning, it can be transformed into an AI Visual Agents that can manipulate smartphones and computers, realizing multi-step complex operations such as sending blessings to designated friends, retouching pictures on computers, and booking tickets on smartphones, and so on.

Related NewsBOCOMI Ratings & TPs on CN Internet Stocks (Table)
For Qwen2.5-1M, Alibaba Cloud’s Tongyi Qianwen has launched two sizes, 7B and 14B, both of which stably outperform GPT-4o-mini in long text processing tasks; meanwhile, the open-source reasoning framework can realize nearly 7 times speedup in processing million-level long text inputs. It is also the first time for the company to extend the context of open source Qwen model to 1M length.