Alibaba’s Qwen3 Becomes Top Open-Source AI Model, Surpassing DeepSeek’s R1
- tech360.tv
- May 7, 2025
- 2 min read
Alibaba Group Holding’s Qwen3 has overtaken DeepSeek’s R1 to become the world’s highest-ranked open-source artificial intelligence model, according to the latest LiveBench benchmark results.
LiveBench, an independent platform that evaluates large language models (LLMs), showed Qwen3 outperforming R1 in coding, mathematics, data analysis and language instruction tasks.
The Qwen3 family, released last week by Alibaba Cloud, includes eight enhanced models ranging from 600 million to 235 billion parameters. Parameters are variables used during AI training to determine how data prompts generate responses.
DeepSeek’s R1 had held the top spot on LiveBench since its launch in January. Qwen3’s rise reflects the rapid development of China’s AI sector and Alibaba’s growing influence in the global open-source community.
Open-source AI models allow public access to source code, enabling developers to modify, share or scale the technology. This approach has played a significant role in China’s tech growth over recent decades.
In February, Alibaba’s earlier Qwen models already powered the world’s top 10 open-source LLMs, according to Hugging Face, a collaborative machine-learning platform.
DeepSeek previously gained global attention by releasing its V3 and R1 models between December 2024 and January, offering high performance at lower costs and computing requirements.
Despite Qwen3’s open-source success, it still trails behind leading closed-source models in overall rankings. OpenAI’s o3-mini high, Google’s Gemini Pro 2.5 and Anthropic’s Claude 3.7 occupy the top three spots.
OpenAI’s o3-mini high costs USD 10 per 1 million tokens to operate. In contrast, Qwen3 costs 4 yuan (USD 0.55) per 1 million tokens, making it significantly more affordable.
Qwen3’s performance and cost advantages have attracted support from major enterprises. Huawei Technologies, Moore Threads, Cambricon Technologies and Hygon Information Technology have all announced support for the model.
Cambricon said last Tuesday it had optimised Qwen3 to run efficiently on its graphics processing units, addressing demand for China-produced chips.
AI infrastructure firms Hyperbolic and Fireworks.ai are adopting Qwen3 on their cloud platforms. Nvidia and Intel have also reportedly begun supporting the model.
Data centres across China, including in Beijing, Shanghai, Hangzhou, Hubei, Jilin and Shaanxi, have adopted Qwen3. The China Supercomputing Network, which links over 20 computing centres in 20 cities across 14 provinces, has also integrated the model.
Alibaba’s Qwen3 is now the top-ranked open-source AI model on LiveBench
The model outperformed DeepSeek’s R1 in coding, maths and data analysis
Qwen3 is more affordable than leading closed-source models like OpenAI’s o3
Source: SCMP
