Huawei Ascend Achieves Zero-Day Adaptation for Alibaba Qwen3-VL-30B-A3B Model, Further Upgrading Multimodal Capabilities

On October 4, the "Huawei Computing" official account officially announced that the Ascend platform has achieved zero-day adaptation for Alibaba Cloud's Tongyi Qianwen Qwen3-VL-30B-A3B series models, providing developers with an immediately usable multimodal AI runtime environment.

The adapted models are Alibaba Cloud's latest open-source Qwen3-VL-30B-A3B-Thinking and Qwen3-VL-30B-A3B-Instruct, both belonging to the Qwen3-VL multimodal visual language model family. This series of models, while inheriting the text processing capabilities of previous generations, focuses on strengthening the visual understanding dimension, bringing eight major core functional upgrades: first, the addition of visual agent capabilities, which can operate PC and mobile interfaces to complete tasks; second, optimized visual encoding, supporting the generation of Draw.io diagrams and HTML/CSS/JS code from audio and video; third, enhanced spatial perception, capable of judging object positions, perspectives, and occlusion relationships, laying the foundation for 3D reasoning; fourth, extended context length to native 256K (expandable to 1M), capable of processing books and hour-long videos with second-level indexing; fifth, strengthened multimodal reasoning, able to output logical evidence and conclusions in STEM and mathematics fields; sixth, upgraded visual recognition, covering celebrities, anime, landmarks, and other categories of objects; seventh, expanded OCR support to 32 languages, optimizing recognition effects in low-light and blurry scenes, and improving handling of rare scripts and specialized terminology; eighth, seamless text-vision fusion, maintaining text understanding levels comparable to pure language models.

In fact, Huawei Ascend has long provided synchronous support for the development of the Qwen series models. After the open-sourcing of the new Qwen3-VL models, they were immediately made out-of-the-box usable in the LLaMA Factory and vLLM frameworks, continuing Ascend's efficient response capabilities in model adaptation.

Previously, on September 29, after the open-sourcing of the DeepSeek-V3.2-Exp model, Huawei completed the adaptation and deployment based on the vLLM/SGLang frameworks that same evening, while synchronously open-sourcing the inference code and operator implementations, once again confirming the Ascend ecosystem's rapid integration capabilities for cutting-edge AI models.