Huawei rose 0day to support Ali Qwen3.
According to the news of IT House, "Huawei Computing" WeChat official account announced yesterday (October 4th) that Shengteng 0day supports Qwen3-VL-30B-A3B model adaptation.
Yesterday, Alibaba Cloud Tongyi Qianwen Qwen3 series model released and opened its new generation of multimodal models: Qwen3-VL-30B-A3B-Thinking and Qwen3-VL-30B-A3B-Instruct. Qwen3-VL is a multi-modal visual language model series, based on its previous generation products,Qwen3-VL has achieved remarkable improvement in visual understanding, while maintaining strong pure text processing ability..
Major enhancements:
Visual agent: operating the graphical user interface of PC/mobile terminal-identifying elements, understanding functions, calling tools and completing tasks.
Visual coding enhancement: generating from images/videos Draw.io/ HTML / CSS / JS。
Advanced spatial perception: judging the position, perspective and occlusion of the object; Provide a stronger 2D foundation and realize a 3D foundation for spatial reasoning and embodied artificial intelligence.
Long context and video understanding: native 256K context, which can be extended to 1 m; Can handle books and hours-long videos, with complete recall and second-level indexing function.
Enhanced Multimodal Reasoning: Excellent in STEM/Mathematics field-able to conduct causal analysis and provide answers based on logic and evidence.
Upgraded visual recognition: Through more extensive and higher-quality pre-training, we can "recognize everything"-celebrities, cartoons, products, landmarks, animals and plants, etc.
Extended OCR function: support 32 languages (previously 19 languages); Robust in low light, blur and tilt; Better handle rare/ancient Chinese characters and technical terms; Improved long document structure parsing.
Text comprehension ability equivalent to pure LLMs: seamless text-visual integration, achieving lossless and unified understanding.
IT House noticed that Ascension had been supporting Qwen series models synchronously before, and this time Qwen3-VL-30B-A3B-Thinking and Qwen3-VL-30B-A3B-Instruct models were released as open source, that is, they were used out of the box in LLaMA Factory and vLLM.Implement 0Day adaptation of the model..
It is worth mentioning that DeepSeek-V3.2-Exp was released and open source on September 29th, introducing sparse Attention architecture. Huawei announced that Shengteng had quickly completed the adaptation deployment based on reasoning frameworks such as vLLM/SGLang.Implement DeepSeek-V3.2-Exp 0day support.And open source all reasoning codes and operator implementations for developers.