Qwen releases QwQ-32B, matching DeepSeek-R1 at 32B parameters

Background

Reasoning models work through a problem step by step before answering. DeepSeek-R1 has 671 billion parameters with 37 billion activated. The Qwen team compared its 32-billion-parameter model with it on the same kinds of tasks.

What happened

The Qwen team released QwQ-32B on 6 March 2025 under the Apache 2.0 license. The team said the model performs comparably to DeepSeek-R1 on math, coding and general problem solving.^[1]

Key details

The training method was reinforcement learning from a cold-start checkpoint. The first stage used an accuracy verifier for math answers and a code-execution server that checks code against test cases. The team chose these tools instead of a traditional reward model.^[1]

A second stage added a general reward model and rule-based verifiers. The team reported gains in instruction following, alignment with human preferences and agent performance, with no significant loss in math and coding.^[1]

The weights were published on Hugging Face and ModelScope. The model could also be used in Qwen Chat and through the DashScope API as qwq-32b.^[1]

Some terms need explaining. A cold-start checkpoint is a model that has received some initial training before reinforcement learning begins, so that reinforcement learning starts from a usable base. Reinforcement learning is a method in which a model is rewarded for the outcomes it produces, such as a correct answer to a math problem.^[1]

A reward model is a separate network trained to score answers, and it is a common way to grade open-ended responses. In this case the team used checkers that follow fixed rules instead. An accuracy verifier compares a math answer with the correct one, and a code-execution server runs a program against test cases. Both give a clear pass or fail, so they can serve as rewards without a scoring network.^[1]

The second stage used rule-based verifiers alongside a general reward model. A model trained this way is checked against both fixed rules and a learned judgement.^[1]

Impact

The comparison put a 32-billion-parameter model beside a model with 671 billion parameters, roughly twenty times larger. The next model family, Qwen3, arrived in April 2025 and combined thinking and non-thinking modes.^[2]