Qwen2.5-VL 7B

Qwen2.5-VL 7B is the 7B instruct variant of Alibaba's Qwen2.5-VL vision-language family, released under Apache 2.0. It takes text, images, and video as input with a 32K-token context window, and built its reputation on document parsing, OCR-heavy workflows, and video understanding — with 5.2M+ Hugging Face downloads, it remains one of the most-deployed open VLMs in production RAG and agent pipelines. Note: it is now the previous generation — Qwen3-VL (already in this catalog) succeeds it with a larger context window and newer architecture. Self-host via vLLM/Transformers/Ollama, or call through Alibaba Cloud Model Studio and third-party hosts.

Provider-reported (Qwen technical report/blog): state-of-the-art among open VLMs at release on document-understanding and video benchmarks. Independent: real-world messy-document tests (2026) found Qwen2.5-VL-class models strong on forms and tables but weaker on non-Latin date formats and very long contracts — treat "SOTA" claims as provider-reported unless you re-test your own workload.