QwQ-32B
Alibaba · China · 2025
The last model Alibaba ever sold as a separate reasoning product — and the one that put open-weights reasoning on a single graphics card.
QwQ-32B was published on 5 March 2025 with open weights under the Apache 2.0 licence. Its claim was arithmetic rather than rhetorical: 32.5 billion parameters, against the 671 billion of DeepSeek-R1, at benchmark scores its authors placed in the same class. Reasoning quality, the release argued, comes from reinforcement learning applied to a solid base model, not from sheer size — and a model of this size fits on hardware an individual can own. Technically it is a Qwen2.5-32B base carried through supervised fine-tuning and reinforcement learning: 64 layers, grouped-query attention with 40 query and 8 key-value heads, and a 131,072-token context window. The manufacturer attaches an unusual caveat to that window — prompts longer than 8,192 tokens require YaRN scaling to be enabled by hand, so the headline figure is not what an unmodified deployment delivers. The line began four months earlier with QwQ-32B-Preview, released on 27 November 2024, and it ended here. Alibaba never shipped a QwQ successor. From Qwen3 onwards, the same company builds thinking into its general models as a switch the caller sets per request, which makes QwQ the closing entry of a product category rather than the first entry of a family. It has not been withdrawn and it is still widely used: the weights are downloaded tens of thousands of times a month, roughly three times as often as the preview that opened the line. An open-weights model retired as a product does not disappear the way a hosted one does — it simply stops gaining successors.
▸News
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!