A unit of Chinese giant Alibaba, Qwen Team, has introduced a new model with 32 billion parameters — QwQ-32B. It provides performance similar to DeepSeek’s R1, but has significantly lower computational requirements, VentureBeat reports.
For example, while DeepSeek-R1 operates with 671 billion parameters, QwQ-32B achieves similar performance at a much smaller size. It typically requires only 24 GB of video memory on the GPU (NVIDIA's H100 has 80 GB), compared to over 1500 GB of video memory for the full version of DeepSeek R1 (16 NVIDIA A100 GPUs), highlighting the efficiency of Qwen's approach.
The model is available in the public domain on Hugging Face and ModelScope under the Apache 2.0 license. This allows for both commercial and research use. Individual users can also access the model via Qwen Chat.
QwQ, or Qwen-with-Questions, was introduced by Alibaba in November 2024 as an open-source model designed to compete with OpenAI's o1.
At launch, the model was designed to improve logical reasoning by reviewing and refining its own answers, making it particularly effective on math and coding tasks. The developers called this process reinforcement learning (RL).
Despite its advantages, early versions of QwQ lagged behind OpenAI in benchmarks, particularly in programming on platforms like LiveCodeBench. In addition, like many new reasoning models, QwQ faced problems such as language mixing and random reasoning loops.
QwQ-32B, thanks to the use of reinforcement learning (RL), can already compete with leading models such as DeepSeek-R1 and o1-mini, despite having fewer parameters.