Alibaba’s Qwen Team has unveiled the QwQ-32B AI model, a new addition to its suite of artificial intelligence frameworks, on Wednesday. This reasoning model utilizes extended test time computation integrated with a visible chain-of-thought (CoT) mechanism. According to the developers, despite its smaller size relative to the DeepSeek-R1, the QwQ-32B performs comparably well on benchmark tests. While the QwQ-32B is classified as an open-source AI model, it is not entirely open for public use.
Release of the QwQ-32B Reasoning AI Model
In a recent blog post, the Qwen Team elaborated on the capabilities of the QwQ-32B reasoning model. The QwQ series, which stands for Qwen with Questions, made its debut in November 2024, aimed at providing an open-source competitor to OpenAI’s o1 series. Featuring 32 billion parameters, the QwQ-32B has been built using advanced reinforcement learning (RL) techniques.
The training methodology employed in developing QwQ-32B included a reinforcement learning scaling approach, starting from a cold-start checkpoint. Initially, RL techniques were restricted to tasks related to coding and mathematics, with careful verification of the output to confirm its accuracy. Over time, this approach was expanded to enhance general capabilities alongside rule-based verifiers, leading to improved performance across various tasks without sacrificing coding and mathematical proficiency.
Benchmarks for the QwQ-32B AI Model
Photo Credit: Alibaba
The Qwen Team asserts that the training structures put in place enable the QwQ-32B to rival the performance of the DeepSeek-R1, which boasts an impressive 671 billion parameters (with 37 billion activated). Internal testing has suggested that the QwQ-32B excels in benchmarks such as LiveBench (coding), IFEval (instruction-based language), and the Berkeley Function Calling Leaderboard V3 (BFCL), which assesses the ability to call functions.
For developers and AI enthusiasts interested in exploring the model, the open weights can be accessed through Hugging Face and Modelscope. The model operates under an Apache 2.0 license, permitting usage for academic and research purposes but prohibiting any commercial applications. However, due to the incomplete training details and datasets, full replication or deconstruction of the model is not feasible. Similarly, DeepSeek-R1 is also distributed under the same licensing terms.
Users who lack the necessary hardware to run the AI model locally can utilize its features through Qwen Chat. The model picker menu located at the top left of the page allows users to select the QwQ-32B preview model.
