QwQ-32B

What QwQ-32B is

QwQ-32B is a reasoning model released by the Qwen team on 6 March 2025 under the Apache 2.0 licence ^[1]. A reasoning model, in the sense the term is used here, is a language model that writes out a long stretch of intermediate thinking before committing to a final answer, instead of answering in a single pass. That extra working lets it grind through multi-step problems, above all in mathematics and code, where a wrong turn early on is hard to recover from later.

The number in the name is the parameter count. The Hugging Face model card gives the figure more precisely: 32.5B parameters, 31.0B of them non-embedding, over 64 layers, with a full context length of 131,072 tokens ^[3].

How it was trained

The team's central argument was about method as much as size. Training started from a cold-start checkpoint and then scaled up reinforcement learning with outcome-based rewards ^[1]. Reinforcement learning, in general terms, means letting the model try, then adjusting it toward the attempts that turned out well; outcome-based rewards means the judgement is made on the final result rather than on each intermediate step.

The reinforcement learning ran in two stages. The first targeted mathematics and coding. Instead of a traditional reward model — a separate network trained to score answers — the team used an accuracy verifier for maths answers and a code-execution server that runs generated code against test cases ^[1]. Both are mechanical checks: an answer is either right or wrong, and code either passes the tests or does not. That removes one source of noise, since a learned reward model can be gamed or can drift away from what people actually want.

The second stage added a general reward model alongside rule-based verifiers. This improved instruction following, alignment with human preferences and agent performance, and the team reported no significant loss in maths and coding ^[1]. In other words, the first stage bought reasoning ability, and the second stage was meant to stop the model from becoming narrow or awkward to talk to.

How it compares

The headline claim is a size comparison: the Qwen team said the 32-billion-parameter model performs comparably to DeepSeek-R1, which has 671B parameters with 37B activated ^[1]. The reported evaluation set covered mathematics, coding and general problem solving, with DeepSeek-R1, DeepSeek-R1-Distilled-Qwen-32B, DeepSeek-R1-Distilled-Llama-70B and o1-mini as the comparison points ^[1]. The Hugging Face model card reports an MMLU-Pro score of 69.07 ^[3], a broad knowledge-and-reasoning benchmark of the kind these comparisons lean on.

Agent capabilities

QwQ-32B also integrates agent capabilities: it can use tools and adapt its reasoning according to feedback from the environment ^[1]. Rather than producing a single chain of thought from memory, it can act, see what came back, and change course. This is why the second training stage cared about agent performance — tool use is only as good as the model's willingness to revise a plan when a tool returns something unexpected.

How to use it

The weights are openly available on Hugging Face and ModelScope, and the model can be used in Qwen Chat or through the DashScope API under the name qwq-32b ^[1]. The model card carries one practical warning: it recommends temperature 0.6 and top-p 0.95 instead of greedy decoding, to avoid endless repetitions ^[3]. A reasoning model that keeps generating is a real failure mode, and slightly random sampling turns out to break the loop where always picking the most likely next token does not.

Where it sits

QwQ-32B belongs to a period when reasoning was a separate model you chose explicitly, rather than a mode inside a general one. Its successor, Qwen3, arrived in April 2025 and merged thinking and non-thinking modes into a single model family ^[2]. Seen from that later vantage point, QwQ-32B is the clearest statement of the earlier idea: that a comparatively small model, trained hard with verifiable rewards, could be talked about in the same breath as models many times its size.