What changed
Alibaba's Qwen team released Apache-2.0 weights for Qwen3-VL-235B-A22B in Instruct and Thinking variants. The model handles images, video, documents, visual grounding and tool-oriented computer interaction, with code and cookbooks.
Why it matters
Teams gained a self-hostable multimodal foundation for document, video and interface workflows rather than relying only on proprietary vision APIs.
What to do
Evaluate the hosted endpoint first on document parsing and grounding. Self-host only if control offsets the multi-GPU serving burden, then pin model, processor and sampling versions.
Caveat
Most benchmark results were Qwen-run. The 235B MoE remains operationally large, and later smaller checkpoints and repository updates should not be attributed to the initial release.