Qwen3-VL-30B-A3B

Qwen3-VL-30B-A3B is the efficiency play in Alibaba's Qwen3-VL series, released September 22, 2025 under Apache 2.0. The mixture-of-experts design holds 30.5B total parameters but activates only about 3.3B per token, delivering near-32B-dense-class visual reasoning at a fraction of the inference cost. It ships in Instruct and Thinking variants with a 256K-token context window, covering spatial understanding, document VQA, video analysis, and GUI agent tasks.

The honest picture: MoE efficiency is real, but routing overhead and the complexity of serving MoE checkpoints mean the practical savings depend heavily on your inference stack; vLLM handles it well, naive transformers serving less so. As the mid-tier of the Qwen3-VL line it sits between the nimble 8B and the flagship 235B MoE, and newer Qwen3.5-VL models have since taken the spotlight. Still, for cost-sensitive multimodal serving it remains one of the best open options.