Alibaba, the Chinese language large, introduced on Thursday a brand new AI mannequin underneath the Qwen umbrella, referred to as QwQ 32B. The mannequin accommodates 32 billion parameters, however is claimed to ‘obtain efficiency comparable’ to DeepSeek-R1, which consists of 671 billion parameters (with 37 billion activated parameters).
The corporate attributes the mannequin’s success to the ‘effectiveness of reinforcement studying (RL)’ when utilized to foundational fashions on a big corpus of information. The QwQ 32B reasoning mannequin additionally carries agentic capabilities, which helps it suppose critically based mostly on exterior suggestions.
“We discover that RL coaching repeatedly improves the efficiency, particularly in math and coding, and we observe that the continual scaling of RL might help a medium-size mannequin obtain aggressive efficiency in opposition to gigantic MoE (combination of specialists) mannequin,” mentioned the corporate in a publish on X.
QwQ-32B is launched as an open weights mannequin and is accessible in Hugging Face and ModelScope. Customers can entry it on-line by way of Qwen Chat.
The mannequin presents efficiency parity with DeepSeek’s flagship R1 mannequin, outperforming OpenAI’s o1-mini in a number of benchmarks pertaining to code, mathematical reasoning, and normal problem-solving duties.
Supply: Qwen
Lately, they launched QwQ-Max-Preview, constructed on the Qwen 2.5 Max, specialising in arithmetic and coding-based duties. On the LiveCodeBench leaderboard, a platform that evaluates LLMs for code, the QwQ-Max-Preview scored 65.6 factors, greater than OpenAI’s o1 medium (63.4) and o3 Mini Low (60.9).
It was then talked about that smaller variants of QwQ reasoning fashions will probably be open sourced for native gadget deployment – and the QwQ 32B mannequin is probably going the primary variant.
Alibaba, too, just lately launched the Wan 2.1, its open-source video basis mannequin, which may generate movies with complicated motions that precisely simulate real-world physics. The suite contains three principal fashions: Wan2.1-I2V-14B, Wan2.1-T2V-14B, and Wan2.1-T2V-1.3B.
The I2V-14B mannequin generates movies at 480P and 720P resolutions, producing complicated visible scenes and movement patterns.
The mannequin outperformed OpenAI’s Sora on the VBench Leaderboard.
Final week, the corporate introduced that it plans to speculate over $52 billion within the cloud computing and synthetic intelligence sector over the following three years. This funding exceeds the corporate’s complete AI and cloud spending previously decade.
In the course of the earnings name, Eddie Wu, CEO of Alibaba Group, mentioned, “We see AI as a once-in-a-generation trade transformation alternative, and the first purpose of our AI technique is to pursue the realisation of AGI (Synthetic Common Intelligence) and repeatedly push the boundaries of mannequin intelligence capabilities.”
The publish Alibaba’s New QwQ 32B Mannequin is as Good as DeepSeek-R1 ; Outperforms OpenAI’s o1-mini appeared first on Analytics India Journal.
