Alibaba Cloud has launched QwQ-32B, a compact reasoning model built on its latest large language model (LLM), Qwen2.5-32B. According to the company, QwQ-32B delivers performance comparable to other large cutting-edge models, including Chinese rival DeepSeek and OpenAI's o1, with only 32 billion parameters.

The model's performance is attributed to the power of reinforcement learning (RL), the core technique behind QwQ-32B, when applied to a robust foundation model like Qwen2.5-32B, which is pre-trained on extensive world knowledge. By leveraging continuous RL scaling, QwQ-32B demonstrates significant improvements in mathematical reasoning and coding proficiency.

Additionally, the model was trained using rewards from a general reward model and rule-based verifiers, enhancing its general capabilities. These include better instruction-following, alignment with human preferences, and improved agent performance. QwQ-32B is open-weight in Hugging Face and Model Scope under the Apache 2.0 license.

Notably, QwQ-32B's 32 billion parameters achieve performance comparable to DeepSeek-R1, which boasts 671 billion parameters (with 37 billion activated). This marks Qwen's initial step in scaling RL to enhance reasoning capabilities. Through this journey, Alibaba has recognized the immense potential of scaled RL and the untapped possibilities within pre-trained language models.

The company is confident that combining stronger foundation models with RL powered by scaled computational resources will propel them closer to achieving Artificial General Intelligence (AGI). Additionally, Alibaba is actively exploring the integration of agents with RL to enable long-horizon reasoning, aiming to unlock greater intelligence with inference time scaling.

Industry experts have weighed in on the launch, with Justin St-Maurice, technical counselor at Info-Tech Research Group, commenting that "comparing these models is like comparing the performance of different teams at NASCAR. Yes, they are fast, but in every lap someone else is winning … so does it matter?" St-Maurice emphasized the importance of aligning models with actual use cases, citing the need for optimization over brute-force algorithms and expensive data centers.

St-Maurice also highlighted the challenges posed by Chinese AI models, including concerns around enterprise risk appetite, regulatory concerns, and data governance policies. He noted that all models challenge ethical boundaries in different ways, and framing one model as inherently more ethical than another is increasingly ambiguous and a matter of opinion.

The launch of QwQ-32B comes as Chinese companies, including Baidu, are making significant strides in AI development. However, St-Maurice pointed out that Baidu's model, Ernie, has had little impact outside of China, citing language barriers and differing priorities as contributing factors.

Overall, Alibaba Cloud's QwQ-32B marks a significant milestone in the development of compact reasoning models, and its performance capabilities are likely to have a significant impact on the AI landscape.