On April 29, Observer Network learned that after Alibaba Qwen3 was open-sourced, the upstream and downstream supply chains conducted overnight adaptation and invocation. Multiple leading chip manufacturers such as NVIDIA, Qualcomm, MediaTek, and AMD have successfully adapted to Qwen3, significantly improving inference efficiency on different hardware platforms and software stacks, meeting AI inference requirements for mobile terminals and data center scenarios.
According to introductions, on authoritative mathematical and code evaluation sets like AIME and LiveCodeBench, the newly open-sourced Qwen3 comprehensively surpasses leading domestic and international models such as DeepSeek-R1 and OpenAl-o1. Qwen3 is also the first hybrid reasoning model in China, capable of "instant response" with low computational power for simple tasks and conducting "in-depth thinking" in multiple steps for complex issues, greatly reducing computational power consumption.
According to official information from Alibaba Cloud, this open-source includes two MoE models: Qwen3-235B-A22B (over 235 billion total parameters, over 22 billion activated parameters), and Qwen3-30B-A3B (30 billion total parameters, 3 billion activated parameters); as well as six Dense models: Qwen3-32B, Qwen3-14B, Qwen3-8B, Qwen3-4B, Qwen3-1.7B, and Qwen3-0.6B.
In addition, just hours after Alibaba Qwen3's open-source, Huawei officially announced support for deploying the full range of Qwen3 models on Ascend, allowing developers to use it immediately in MindSpore and MindIR, achieving zero-day adaptation for Qwen3. Subsequently, Hygon Information also stated that under the guidance of the "DeepCompute Intelligence" strategy, Hygon DCU quickly completed seamless adaptation and optimization for all eight models, covering 235B/32B/30B/14B/8B/4B/1.7B/0.6B, achieving second-level deployment without errors or compatibility issues.
As early as March 2024, smartphone chip manufacturer MediaTek had already deployed Tongyi Qwen18B and 40B parameter large models on flagship chips such as Dimensity 9300. This adaptation of Qwen3 is a continuation of the chip-level soft and hard adaptation between both parties. Last October, Qualcomm announced cooperation with Tencent Hunyuan and Zhipu during its product launch, where Tencent Hunyuan large models in versions 7B and 3B and Zhipu's GLM-4V end-side visual large model will be available on phones equipped with Snapdragon 8 Elite.
This article is an exclusive contribution by Observer Network and cannot be reproduced without permission.
Original Source: https://www.toutiao.com/article/7498652333929447970/
Disclaimer: The views expressed in this article are those of the author alone. Please express your opinions by clicking the "Agree/Disagree" buttons below.