Qwen2.5-VL
Vision-language models tuned as visual agents, from 3B to 72B parameters.
About
Released in January 2025, Qwen2.5-VL builds on Qwen2-VL with stronger parsing of diagrams, charts, and layouts plus hour-long video understanding, pitched explicitly as a visual agent: Alibaba demoed it booking flights and navigating GIMP from screen content. Team-reported VLMEvalKit numbers for the 72B: 70.0 MMMU, 96.1 DocVQA, 88.2 MMBench, 75.3 MathVista; the 3B reportedly matches the previous generation's 7B. The license split matters here: the 3B is research-only (no commercial use), the 7B is Apache 2.0, and the 72B carries the Qwen License with a 100M monthly-active-user commercial cap. Honest caveat: the agent demos are controlled showcases — real UI automation still needs substantial scaffolding around the model.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported (VLMEvalKit): 72B — 70.0 MMMU, 96.1 DocVQA, 88.2 MMBench, 75.3 MathVista.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: January 2025 release verified via HF citation and press; exact day not pinned so releasedAt left null.; License split verified (3B research-only, 7B Apache 2.0, 72B Qwen License with 100M MAU cap) via press coverage.; Team-reported VLMEvalKit numbers (72B: 70.0 MMMU, 96.1 DocVQA) labeled as team-reported.; Agent-demo honesty caveats included..