QwQ-32B-Preview is an experimental research model from the Qwen Team, announced in a blog post dated November 28, 2024, and focused on AI reasoning. The Hugging Face card lists about 32.5B parameters, a 32,768-token context length and the Apache-2.0 license, and says it is built on Qwen2.5-32B. It was a preview, and on March 6, 2025 the Qwen Team published its QwQ-32B reasoning model, trained with scaled reinforcement learning and also released under Apache 2.0.
Key features
- Architecture: per the card, a causal language model with RoPE, SwiGLU, RMSNorm and attention QKV bias, 64 layers, and grouped-query attention with 40 query heads and 8 key-value heads.
- Context: 32,768 tokens.
- Vendor benchmark claims from the blog post: GPQA 65.2%, AIME 50.0%, MATH-500 90.6% and LiveCodeBench 50.0%. These are Qwen's own results.
- Stated limitations: unexpected language mixing or switching, circular reasoning loops that produce long answers without a conclusion, a need for stronger safety measures, and weaker common-sense reasoning and nuanced language understanding than its math and coding strength.
- License: Apache-2.0 on the Hugging Face card.
Use cases
- Studying early open reasoning models and how their outputs differ from standard chat models.
- Math and coding experiments on local hardware that can host a 32B model.
- Comparing the preview with the later QwQ-32B release.
Pricing and availability
The weights are free to download from Hugging Face under Apache-2.0. Qwen describes the model as an experimental research model, so treat it as a preview, not a production default. The card requires a recent transformers version, since versions before 4.37.0 fail with a KeyError: 'qwen2'. We checked the sources on 2026-10-29. The cards list no hosted price.
Getting started
- Update
transformersto a recent version. - Load
Qwen/QwQ-32B-Previewfrom Hugging Face and use its chat template. - Allow long generations, since reasoning outputs can be lengthy, and watch for language mixing or loops.
- For new work, evaluate QwQ-32B (March 2025), which Qwen released as a follow-up reasoning model.
FAQ
What is QwQ-32B-Preview?
An experimental 32B reasoning model from the Qwen Team, announced November 28, 2024.
What is its context length?
32,768 tokens, per the Hugging Face card.
What came after it?
QwQ-32B, published March 6, 2025 under Apache 2.0, trained with scaled reinforcement learning.
What are its known problems?
Qwen lists language mixing, circular reasoning loops, a need for stronger safety measures, and weaker common-sense reasoning.
Is it open source?
The weights are on Hugging Face under Apache-2.0.
Alternatives
- DeepSeek-R1: a reasoning model from DeepSeek; Qwen's QwQ-32B post compares against it.
- Qwen2.5 72B Instruct: a larger general instruct model from Qwen.
- Qwen2.5 Coder 32B Instruct: a 32B coding model from Qwen.
Conclusion
QwQ-32B-Preview is an early open reasoning model with clear limits and an open license, now followed by QwQ-32B. Read the blog post and model card, and browse more in the models category.
Sources
Comments
No comments yet. Be the first to comment!
Related Tools
Related Insights
Point Codex at any model with magpie
Install magpie, add a local or signed-in provider, then switch Codex to that model. Rechecked against official magpie docs on 2026-10-06.
How to pick an AI coding CLI for your repo
Match Claude Code, Codex CLI, OpenCode, Pi, omp, DeepSeek Harness, and Grok Build to the work you actually run. Docs rechecked 2026-10-06.
Running low on ChatGPT Codex quota? Switch to DeepSeek or Grok inside Codex
Codex Router lets you keep your Codex workspace while using DeepSeek, Grok, and other external models. A plain-English guide to routing, quotas, and network access.