Alibaba **Qwen (Tongyi Qianwen)** team recently released two new small-scale multimodal models: Qwen3-VL-30B-A3B-Instruct and Qwen3-VL-30B-A3B-Thinking. Each model has 3 billion active parameters, but it has ambitious performance goals.
According to the Qwen team, these two new versions aim to compete with top models such as GPT-5-Mini and Claude4Sonnet, and have shown stronger performance in key areas such as mathematics, image recognition, text recognition, video processing, and agent control in certain benchmark tests.
This product series also includes an FP8 version for faster inference, as well as an FP8 variant of the Qwen3-VL-235B-A22B model.
These models are now fully available for access, and users and developers can download and use them through HuggingFace, ModelScope, and GitHub, or call them via Alibaba Cloud API. In addition, Qwen also provides a web chat interface that can be used directly. This release highlights the efforts of the Qwen team in promoting the popularization of efficient and high-performance small models.
