Alibaba’s counter to DeepSeek is Qwen 2.5-Max, the organization’s most recent Mixture-of-Experts (MoE) expansive model.
Qwen 2.5-Max features pretraining on over 20 trillion tokens and fine-tuning through state-of-the-art methodologies like Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF).
With the API now accessible via Alibaba Cloud and the model available for experimentation through Qwen Chat, the Chinese tech powerhouse is encouraging developers and researchers to witness its advancements in action.
Surpassing competitors
When evaluating Qwen 2.5-Max’s effectiveness against several top AI models across a range of benchmarks, the findings are encouraging.
Analyses included well-known metrics such as the MMLU-Pro for university-level problem resolution, LiveCodeBench for programming proficiency, LiveBench for overall skills, and Arena-Hard for evaluating models against human preferences.
According to Alibaba, “Qwen 2.5-Max surpasses DeepSeek V3 in benchmarks like Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond, while also showing competitive outcomes in other evaluations, including MMLU-Pro.”
The instruct model – tailored for downstream activities like chat and coding – competes directly with top models such as GPT-4o, Claude-3.5-Sonnet, and DeepSeek V3. Among these, Qwen 2.5-Max was able to outdo competitors in several critical areas.
Comparative analysis of base models also yielded favorable results. Although proprietary models like GPT-4o and Claude-3.5-Sonnet were hard to reach due to access limitations, Qwen 2.5-Max was evaluated against leading public counterparts like DeepSeek V3, Llama-3.1-405B (the largest open-weight dense model), and Qwen2.5-72B. Once again, Alibaba’s new entrant exhibited outstanding performance consistently.
“Our base models have revealed considerable advantages across most benchmarks,” Alibaba asserted, “and we are hopeful that enhancements in post-training methodologies will elevate the forthcoming version of Qwen 2.5-Max to unprecedented heights.”
Making Qwen 2.5-Max available
To enhance the model’s availability to the worldwide community, Alibaba has integrated Qwen 2.5-Max with its Qwen Chat platform, allowing users to interact directly with the model in various roles—be it exploring its search functionalities or evaluating its comprehension of complex inquiries.
For developers, the Qwen 2.5-Max API is currently offered through Alibaba Cloud under the model designation “qwen-max-2025-01-25”. Interested parties can begin by registering for an Alibaba Cloud account, enabling the Model Studio service, and creating an API key.
The API is even compatible with OpenAI’s ecosystem, facilitating easy integration for existing projects and workflows. This compatibility reduces obstacles for those looking to test their applications with the model’s abilities.
Alibaba has made a strong declaration of intent with Qwen 2.5-Max. The organization’s sustained commitment to expanding AI models is not solely about enhancing performance metrics but also about strengthening the core reasoning and cognitive skills of these systems.
“The growth of data and model size not only highlights advancements in model intelligence but also signifies our steadfast dedication to pioneering research,” Alibaba emphasized.
Looking toward the future, the team plans to extend the limits of reinforcement learning to cultivate even more sophisticated reasoning capabilities. They believe this could allow their models to not only equate but exceed human intelligence in tackling intricate issues.
The ramifications for the industry could be substantial. As scaling strategies improve and Qwen models penetrate new territories, we may witness further ripples across AI-driven sectors worldwide, similar to what we’ve seen in recent weeks.
(Photo by Maico Amorim)
See also: ChatGPT Gov aims to modernise US government agencies
Want to discover more about AI and big data from industry frontrunners? Check out the AI & Big Data Expo occurring in Amsterdam, California, and London. The extensive event is co-located with other leading conferences including Intelligent Automation Conference, BlockX, Digital Transformation Week, and Cyber Security & Cloud Expo.
Explore additional upcoming enterprise technology events and webinars powered by TechForge here.
Be the first to comment