QwQ-32B-Preview

What QwQ-32B-Preview is

QwQ-32B-Preview is an experimental reasoning research model released by the Qwen team on 28 November 2024^[1]. It has 32 billion parameters and a context length of 32K tokens, and its weights were published under the Apache 2.0 licence^[2]. The name marks it as a preview: the team presented the release as a research artefact rather than a finished product, and shipped it with an explicit list of things it does badly.

A reasoning model, in the sense used here, is one trained to work through a problem in a long chain of intermediate steps before answering — writing out attempts, checking them, sometimes backing up — instead of producing the answer directly. That approach tends to pay off where the path to the answer is not obvious: multi-step arithmetic, competition mathematics, and code that has to satisfy several constraints at once.

The reported scores

The Qwen team published four benchmark results for the model: GPQA 65.2%, AIME 50.0%, MATH-500 90.6% and LiveCodeBench 50.0%^[1]. The Hugging Face model card also reports an MMLU-Pro score of 70.97^[4]. The four benchmarks probe different things: GPQA is a set of graduate-level science questions written to resist simple look-up, AIME is a competition mathematics exam, MATH-500 is a collection of mathematics problems, and LiveCodeBench tests code generation. Read together, the numbers describe a model that is much stronger on mathematics than on code generation, and that gets about half of AIME-level problems right.

Specifications and how to get it

The Hugging Face model card lists 32.5B total parameters, of which 31.0B are non-embedding parameters^[4]. The full context length given there is 32,768 tokens, under the Apache-2.0 licence, alongside the MMLU-Pro figure^[4]. The Qwen team made QwQ-32B-Preview available through a Hugging Face model collection and demo Space, on ModelScope, and in Qwen Chat^[1].

Limitations, in the team's own list

The Qwen team listed several limitations for the model: it mixes and switches between languages unexpectedly, it can fall into circular reasoning that produces long answers without ever reaching a conclusion, and it needs stronger safety measures^[1]. The team also said the model was strong in mathematics and coding but weaker in common-sense reasoning and nuanced language understanding^[1]. Users were told to be cautious when deploying it, and that it needed enhanced safety measures for reliable and secure performance^[1].

These are recognisable failure modes for a model trained to think out loud at length. The extra tokens help on hard problems, but nothing in the training objective guarantees that the thinking terminates, stays in one language, or remains on topic. Publishing the list alongside the benchmark table was part of how the release was framed: a preview whose edges are visible is easier to build on than one presented as complete.

Where it sits

QwQ-32B-Preview was followed in December 2024 by QvQ-72B, a preview aimed at visual reasoning^[2]. It was succeeded by QwQ-32B, released on 6 March 2025 under Apache 2.0^[3]. The preview therefore sits at the start of a short line of releases: an experimental reasoning model at 32B scale, a larger visual counterpart, and then a successor released as open weights.

Its place in the story is as a public, downloadable example of the step-by-step reasoning approach at a size that could be run outside a closed service, released under a permissive licence with its weaknesses documented as plainly as its scores^[2][1]. That combination — open weights, published benchmark numbers, and an honest list of failure modes — is what makes it useful to look back on.