Qwen 2.5-Max outperforms DeepSeek V3 in some benchmarks
Alibaba’s response to DeepSeek is Qwen 2.5-Max, the company’s latest Mixture-of-Experts (MoE) large-scale model.
Qwen 2.5-Max boasts pretraining on over 20 trillion tokens and fine-tuning through cutting-edge techniques like Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF).
With the API now available through Alibaba Cloud and the model accessible for exploration via Qwen Chat, the Chinese tech giant is inviting developers and researchers to see its breakthroughs firsthand.
Outperforming peers
When comparing Qwen 2.5-Max’s performance against some of the most prominent AI models on a variety of benchmarks, the results are promising.
Evaluations included popular metrics like the MMLU-Pro for college-level problem-solving, LiveCodeBench for coding expertise, LiveBench for overall capabilities, and Arena-Hard for assessing models against human preferences.
According to Alibaba, “Qwen 2.5-Max outperforms DeepSeek V3 in benchmarks such as Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond, while also demonstrating competitive results in other assessments, including MMLU-Pro.”
The instruct model – designed for downstream tasks like chat and coding – competes directly with leading models such as GPT-4o, Claude-3.5-Sonnet, and DeepSeek V3. Among these, Qwen 2.5-Max managed to outperform rivals in several key areas.
Comparisons of base models also yielded promising outcomes. While proprietary models like GPT-4o and Claude-3.5-Sonnet remained out of reach due to access restrictions, Qwen 2.5-Max was assessed against leading public options such as DeepSeek V3, Llama-3.1-405B (the largest open-weight dense model), and Qwen2.5-72B. Again, Alibaba’s newcomer demonstrated exceptional performance across the board.
“Our base models have demonstrated significant advantages across most benchmarks,” Alibaba stated, “and we are optimistic that advancements in post-training techniques will elevate the next version of Qwen 2.5-Max to new heights.”
Making Qwen 2.5-Max accessible
To make the model more accessible to the global community, Alibaba has integrated Qwen 2.5-Max with its Qwen Chat platform, where users can interact directly with the model in various capacities—whether exploring its search capabilities or testing its understanding of complex queries.
For developers, the Qwen 2.5-Max API is now available through Alibaba Cloud under the model name “qwen-max-2025-01-25”. Interested users can get started by registering an Alibaba Cloud account, activating the Model Studio service, and generating an API key.
The API is even compatible with OpenAI’s ecosystem, making integration straightforward for existing projects and workflows. This compatibility lowers the barrier for those eager to test their applications with the model’s capabilities.
Alibaba has made a strong statement of intent with Qwen 2.5-Max. The company’s ongoing commitment to scaling AI models is not just about improving performance benchmarks but also about enhancing the fundamental thinking and reasoning abilities of these systems.
“The scaling of data and model size not only showcases advancements in model intelligence but also reflects our unwavering commitment to pioneering research,” Alibaba noted.
Looking ahead, the team aims to push the boundaries of reinforcement learning to foster even more advanced reasoning skills. This, they say, could enable their models to not only match but surpass human intelligence in solving intricate problems.
The implications for the industry could be profound. As scaling methods improve and Qwen models break new ground, we are likely to see further ripples across AI-driven fields globally that we’ve seen in recent weeks.
(Photo by Maico Amorim)
See also: ChatGPT Gov aims to modernise US government agencies
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is co-located with other leading events including Intelligent Automation Conference, BlockX, Digital Transformation Week, and Cyber Security & Cloud Expo.
Explore other upcoming enterprise technology events and webinars powered by TechForge here.
