Alibaba releases Qwen2.5-VL 72B: best open-source multimodal model beats GPT-4o on key benchmarks

Alibaba has released Qwen2.5-VL 72B, a powerful AI model that can see and understand images, videos, documents, and application screenshots. The big deal here is that it is fully open source under the Apache 2.0 license — meaning anyone can download, use, and build on top of it for free.

What can it actually do? A lot. It reads scanned documents in many languages, interprets charts and tables, analyzes videos up to 20 minutes long, and can even interact with graphical interfaces the way a human would — identifying buttons, filling in fields, and navigating screens.

The most striking result is the benchmark comparison: on several standard industry tests, Qwen2.5-VL 72B outperforms OpenAI's GPT-4o, which had long been considered the gold standard for multimodal AI.

For developers using tools like Ollama on their own machines, the model is available to run locally — though the 72 billion parameters require serious hardware, typically one or more GPUs with 40 to 80 GB of VRAM combined.

This release marks a turning point for open-source AI. For the first time, a free and open model can genuinely compete with — and in some areas surpass — the best paid offerings from major American tech companies. It shifts the calculus for any organization deciding whether to use cloud AI APIs or run their own models.