Alibaba Open-Sources Its Latest Reasoning Model: QwQ-32B, Comparable to DeepSeek-R1 195

In the early hours of March 6, 2025, Alibaba officially released and open-sourced their latest reasoning model — Tongyi Qianwen QwQ-32B.

Tongyi Qianwen QwQ-32B is trained using large-scale reinforcement learning. Unlike previous pre-training and post-training methods, reinforcement learning enables the model to acquire knowledge through interaction with its environment. As a result, it excels particularly in mathematical reasoning, coding, and general capabilities, allowing for more effective improvements in reasoning ability.

In addition, the model incorporates capabilities related to Agents. With these capabilities, the model can engage in critical thinking when using tools, and dynamically adjust its reasoning process based on feedback from the environment.

Overview of Tongyi Qianwen QwQ-32B Model

• Tongyi Qianwen QwQ-32B is a reasoning model with 32 billion parameters. Although its parameter count is significantly smaller than that of DeepSeek-R1 (which has 671 billion parameters, of which 37 billion are activated), its performance is comparable to the full-blooded version of DeepSeek-R1.

• This model adopts the Apache 2.0 open-source license, allowing users to download it for free and use it for commercial purposes. There are two open-source platforms for the model: Hugging Face (https://huggingface.co/Qwen/QwQ-32B ) and ModelScope (https://modelscope.cn/models/Qwen/QwQ-32B ).

Performance of Tongyi Qianwen QwQ-32B Model

• Outstanding Performance: QwQ-32B performs remarkably well in mathematical reasoning, coding, and general capabilities. In the AIME24 evaluation set and LiveCodeBench tests, its performance is on par with DeepSeek-R1, outperforming other models of the same size.

• Reinforcement Learning and Agent Assistance: Through large-scale reinforcement learning, QwQ-32B has achieved significant improvements in reasoning capabilities, enabling it to think deeply and handle complex problems. Moreover, the model integrates agent capabilities, allowing it to engage in critical thinking when using tools and adjust its reasoning process based on environmental feedback.

• Adaptable Across Multiple Domains: QwQ-32B performs excellently in mathematics, coding, and general task processing, and also supports multimodal tasks such as speech transcription and image recognition with ease. In benchmarks such as LiveBench, IFEval, and BFCL, its scores surpass those of DeepSeek-R1. Additionally, it has low hardware resource requirements, allowing for local deployment on consumer-grade GPUs, making it well-suited for application scenarios with limited hardware resources.

Tongyi Qianwen QwQ-32BModel Applications

• Mathematical and Logical Reasoning: In mathematical reasoning tasks such as AIME24, QwQ-32B achieves excellent results and can be applied in both education and scientific research fields.

• Coding and Code Generation: In LiveCodeBench tests, it generates high-quality code, making it suitable for software development and coding education.

• General Task Processing: QwQ-32B performs exceptionally well in instruction following (IFEval), multimodal task processing, and more, capable of handling a variety of general tasks.

• Agent-Related Tasks: Because it supports agent capabilities, it can engage in dynamic decision-making and adapt to environments, making it suitable for the automated handling of complex tasks.

HF: https://huggingface.co/Qwen/QwQ-32B

Model: https://modelscope.cn/models/Qwen/QwQ-32B

Demo: https://huggingface.co/spaces/Qwen/QwQ-32B-Demo

Chat: https://chat.qwen.ai