Alibaba launches QwQ, its open step-by-step reasoning model

Alibaba's Qwen team has unveiled QwQ-32B-Preview, a roughly 32-billion-parameter language model designed to solve complex math and coding problems by reasoning step by step before producing an answer. The company has released it as an open model, with weights available for download, letting developers and researchers run and study it without relying on a paid API.

What a reasoning model is

Most large language models generate their answer in one pass, word by word, with no explicit deliberation beforehand. So-called reasoning models work differently: before answering, they produce a chain of intermediate steps—a kind of internal scratchpad—in which they break the problem down, try different approaches, and catch their own mistakes before settling on a final answer. That ability markedly improves performance on tasks requiring chained logic, such as mathematical proofs or debugging code, though it comes at a cost: thinking through more steps means taking longer and burning more compute per response.

OpenAI had opened this path in September with o1-preview and o1-mini, two models that showed clear gains on math and science benchmarks compared with GPT-4o, but which the company kept closed: accessible only through its API, with no disclosure of the internal reasoning they generate.

Alibaba's bet: open up the model

QwQ-32B-Preview, a name the company hasn't officially explained, builds on Alibaba's Qwen2.5 family and focuses specifically on this extended reasoning capability. The company has published it on Hugging Face under an open license, meaning anyone with enough hardware can download it, run it locally, and modify it—something no model in OpenAI's o1 family allows.

The name itself carries the "Preview" label: Alibaba is presenting the model as a preliminary version, not a finished product, and warns of limitations typical of a system still under development, such as possible reasoning loops or unexpected language switches mid-response. At 32 billion parameters, it's a mid-to-large-sized model, far from the scale of the biggest closed models but big enough to hold its own on math and coding benchmarks.

Why it matters

Until now, o1-style step-by-step reasoning had been almost exclusively OpenAI's turf. Alibaba releasing an open model aimed at the same problem changes the picture: it narrows the gap between what was only available through a closed, paid API and what any lab, university, or company can download and adapt at zero licensing cost. It also confirms that the "think before you answer" approach isn't a proprietary technique belonging to a single lab, but a research direction that several teams—including Chinese ones—are pursuing in parallel.

For the non-technical user, the practical difference is that this type of model tends to make fewer mistakes on problems requiring several chained logical steps—an equation with multiple unknowns, a function with several conditions—though at the cost of slower responses. For the industry, QwQ-32B-Preview is one more sign that the race for reasoning models is now being fought on two fronts: the closed systems built by major U.S. labs, and the open alternatives that Chinese companies like Alibaba are willing to put in anyone's hands.