QwQ-32B-Preview

QwQ-32B-Preview put weights for a chain-of-thought reasoning model on Hugging Face on 27 November 2024, with the Qwen team's write-up following a day later. At that point reasoning was something you rented: OpenAI had shown o1-preview in September 2024 and kept its thinking hidden. This release let anyone download a 32.5-billion-parameter model, watch every step of its reasoning and fine-tune it, under a plain Apache 2.0 licence with no additional clauses. The scores the team published were narrow and high: 90.6 per cent on MATH-500, 65.2 per cent on GPQA, 50.0 per cent on AIME and 50.0 per cent on LiveCodeBench. Mathematics and code, in other words, and the release says as much — it states outright that the model has room for improvement in common-sense reasoning and nuanced language understanding. What makes the release unusual is the honesty of its limitations section, which is worth reading as a product decision rather than boilerplate. The authors list four defects: the model mixes languages and switches between them mid-answer; it falls into circular reasoning and produces long replies that never reach a conclusion; it needs stronger safety measures before deployment; and its strength is confined to technical domains. Publishing a model in that state, described in those terms, is what a research preview is supposed to mean. The body is a Qwen2.5-32B-Instruct carried through further training: 64 layers, grouped-query attention with 40 query heads against 8 key/value heads, and a 32,768-token context — a quarter of what DeepSeek-R1 would offer two months later. The successor, QwQ-32B, arrived on 5 March 2025 on the same parameter count and closed most of these gaps; the preview stays on Hugging Face as a download and as the marker of where the open reasoning line began.

▸Videos

No videos yet.

▸Reviews