Alibaba has just unveiled its latest reasoning model, and it seems that DeepSeek and OpenAI might have something to worry about — at least if all of Alibaba’s promises turn out to be true. It’s open-source, so I checked it out. You can try it out for free, too, although unsurprisingly, you’ll find that there are some things it won’t talk to you about.
The new model, dubbed QwQ-32b (Quan-with-Questions) runs on much fewer parameters, meaning that it requires less resources, but Alibaba claims that it performs at the same level as DeepSeek or OpenAI’s o1-mini.
DeepSeek’s R1 large language model (LLM) was all the rage earlier in February when it came out, suddenly capable of rivaling the golden standard set by ChatGPT and other alternatives, but at a much lower cost. It seems that Alibaba might be pushing the envelope even further here.
As explained by VentureBeat, DeepSeek-R1 requires 671 billion parameters to run, 37 billion of which are activated. Meanwhile, Alibaba’s new QwQ-32b can get by with 32 billion parameters. Those numbers are totally abstract to many, but there’s a huge difference in compute power; while DeepSeek R1 requires 1600GB of VRAM to run, QwQ-32b can get by with just 24GB of VRAM. In most cases, this will mean Nvidia’s H100 or equivalents, but even the gaming-focused RTX 4090 sports 24GB. The latest RTX 5090 ups that to 32GB.
Alibaba’s QwQ-32b is available under an Apache 2.0 license, meaning that companies and researchers can use it. More importantly, we can use it by trying out Alibaba’s Qwen Chat. Like DeepSeek, it comes with some limitations, but also has a couple of immediate perks that I noticed quite quickly.
It seems to give quite in-depth answers even to quick, simple questions. This can be good, but in a way, it was mostly annoying as it gives you a lot of unnecessary context that you didn’t ask for. I like that it shows you its whole reasoning process, though, which is similar to ChatGPT’s Deep Thinking feature — but with much less depth.
When asked about political matters, Qwen Chat flags it as inappropriate. There might be ways to jailbreak it — it was possible with DeepSeek, after all — but I haven’t managed to just yet.
Whether Alibaba’s claims turn out to be true remains to be seen, but it looks like ChatGPT and DeepSeek now have a new rival.