Qwen releases the open reasoning model QwQ-32B-Preview

What a reasoning model is

A reasoning model is a language model that works through intermediate steps — often a long stretch of text — before it produces a final answer. The point is that spending more computation on the way to an answer can help with problems where a single pass is not enough: multi-step arithmetic, competition-style mathematics, or a piece of code whose behaviour has to be traced through several branches. QwQ-32B-Preview belongs to this category; the Qwen team described it as an experimental research model for reasoning ^[1].

The release

The Qwen team published QwQ-32B-Preview on 28 November 2024 ^[1]. The model was described as a 32B-parameter system with a 32K-token context length, with its weights released under the Apache 2.0 licence ^[2]. The Hugging Face model card gives more precise figures: 32.5 billion total parameters and a full context length of 32,768 tokens ^[3]. The team said the model was made available through Hugging Face, ModelScope and Qwen Chat ^[1].

Releasing the weights under a permissive licence means other people can download the model and run it themselves, rather than only sending queries to a hosted service ^[2]. The "preview" in the name and the "experimental research" framing both signal that this is a working step in a research direction rather than a finished product ^[1].

Reported results

The team reported scores on four benchmarks: 65.2% on GPQA, 50.0% on AIME, 90.6% on MATH-500 and 50.0% on LiveCodeBench ^[1]. These tests probe different things — GPQA consists of graduate-level science questions written to resist simple lookup, AIME is a competition mathematics exam, MATH-500 is a set of mathematics problems, and LiveCodeBench evaluates code generation. Read together, the numbers show where the model was strongest and where it was not.

The team's own summary matches that reading: the model was strong in mathematics and coding but weaker in common-sense reasoning and nuanced language understanding ^[1].

Stated limitations

The Qwen team listed several problems with the release ^[1]. One is unexpected language mixing and switching, where an answer drifts between languages instead of staying in the language of the question. Another is circular reasoning: the model can produce long answers that never arrive at a conclusion. The team also said stronger safety measures were needed ^[1].

These are the kinds of failure that come with the territory of models that generate long chains of intermediate text: the longer the working, the more chances for it to loop, wander, or change register. Publishing the list alongside the benchmark numbers made the trade-off visible to anyone who wanted to try the model.

What it represented

QwQ-32B-Preview put a reasoning-oriented model into the open, at a size that fits on modest hardware compared with the largest systems, under a licence that allows reuse ^[2]. Combined with the explicit list of weaknesses and the benchmark results, the release gave outside researchers a concrete object to study — a model whose strengths in mathematics and coding, and whose failures in common sense, language consistency and looping, could be examined directly rather than inferred from a leaderboard ^[1].