Alibaba has launched its newest reasoning model and it could pose a challenge to competitors like DeepSeek and OpenAI—assuming Alibaba’s claims hold up. The model is open-source, so I decided to take a look for myself. You can try it out for free as well, but expect certain topics to be off-limits.
The model, named QwQ-32b (Quan-with-Questions), operates with significantly fewer parameters, which means it requires less computing power. Nevertheless, Alibaba asserts that it performs comparably to DeepSeek’s and OpenAI’s o1-mini models.
In February, DeepSeek’s R1 large language model (LLM) gained attention for its ability to compete with the high standards set by ChatGPT and other alternatives at a lower cost. It appears that Alibaba is taking strides to push the boundaries even further.
According to VentureBeat, DeepSeek-R1 requires a staggering 671 billion parameters to function, with 37 billion being active at any given time. In contrast, Alibaba’s QwQ-32b only needs 32 billion parameters. While these figures might seem abstract, the difference in computational needs is significant; DeepSeek R1 demands 1600GB of VRAM whereas QwQ-32b can operate with just 24GB. This means that while DeepSeek typically requires powerful systems like Nvidia’s H100, the gaming-oriented RTX 4090 has the necessary 24GB, and the newer RTX 5090 offers 32GB.
Alibaba’s QwQ-32b is released under an Apache 2.0 license, allowing companies and researchers, as well as everyday users, to experiment with it via Alibaba’s Qwen Chat. Though it also has its limitations, I noticed some immediate advantages.
It appears to provide comprehensive answers to even the most straightforward questions. While this depth of response can be beneficial, it can sometimes overwhelm with superfluous information. I appreciate that it reveals its reasoning process, similar to features found in ChatGPT, though it’s less detailed.
When it comes to discussing political topics, Qwen Chat marks them as inappropriate. Although bypassing this limitation may be possible—similar to hacks used with DeepSeek—I haven’t achieved that yet.
Whether Alibaba’s assertions will hold true remains to be seen, but it seems that ChatGPT and DeepSeek now face new competition.
