QwQ-32B

QwQ-32B is a language model designed for reasoning and dialogue in the Qwen series. It is a medium-sized reasoning model with capabilities aimed at competitive performance on hard problems. It is intended for use via a model hub and related interfaces, including demos and chat services, within the Hugging Face ecosystem.

The Quickstart section shows how to load the tokenizer and model using the transformers library, illustrating the typical workflow for using QwQ-32B. It includes example code that demonstrates loading the pretrained model, tokenizer, and generating content from a prompt. It emphasizes the integration with the transformers framework and how prompts are prepared for generation.

The model’s usage guidelines specify recommended decoding settings and prompt conventions to standardize outputs in benchmarking. It discusses handling long inputs with YaRN for extended context and describes deployment recommendations with vLLM, along with notes about output formatting and multi-turn history. These guidelines are meant to help users optimize performance and output quality.

QwQ-32B’s documentation mentions its size, architecture, and technical details such as parameter counts, layer count, attention heads, and the 131,072-token context length. It notes that YaRN should be enabled for long prompts and highlights the availability of demonstrations and access through QwenChat. It also points to further details in the blog, GitHub, and Documentation for more information.

Use cases

• Reasoning tasks and complex prompts