Alibaba’s Quven Team, a division assigned the work of developing the Artificial Intelligence (AI) model, released the QWQ-32B AI model on Wednesday. It is an argument model based on extended testing time calculations with visual chain-key-thinking (COT). Developers claim that despite being smaller in size than Deepsek-R1, the model can match its performance based on benchmark score. Like other AI models released by the Quven team, QWQ-32B is also an open-source AI model, however, it is not completely open.
QWQ-32B Rogue AI model released
One in blog postAlibaba’s Quven team expanded the QWQ-32B Reasoning model. QWQ (small for Qwen with questions) series AI models were first introduced by the company in November 2024. These logic models were designed to offer an open-source option to the choice of O1 series of Openai. QWQ-32B is a 32 billion parameter model developed by scaling reinforcement learning (RL) techniques.
Explaining the training process, the developers stated that the RL scaling approach was added to the cold-start checkpoint. Initially, RL was used only for coding and mathematics related functions, and the reactions were verified to ensure accuracy. Later technology was used for general capabilities with rules-based verifiers. The Quven team found that this method increased the general capabilities of the model without reducing its mathematics and coding performance.
QWQ-32B AI Model Benchmark
Photo Credit: Alibaba
Developers claim that these training structures enabled the QWQ-32B to perform at the same levels of the Deepsek-R1, despite being a 671-Billion-Permitter model (with 37 billion active). Based on the internal test, the team claimed that the QWQ-32B Livebench (Coding), IFEVAL (chat or Instruction Fine-Tund Language), and the Berkeley Function Calling Calling Leaderboard V3 or BFCL (Call’s Call Call) performs better than the QWQ-32B Deepsac-R1 in benchmark.
The enthusiasts of developers and AI can get the open weight of the model when they hug face listing and modelcope. The model Apache 2.0 is available under the license that allows academic and research related use but forbids cases of commercial use. Additionally, since complete training details and datasets are not available, the model is also not replica or it can be decontrolled. The Deepsek-R1 was also available under the same license.
If someone has the lack of perfect hardware to run the AI model locally, they can also use its abilities through Quven Chat. The model picker menu at the top-left of the page will allow users to select the QWQ-32B-Preview model.
