ai/qwq repository overview

⁠QwQ

⁠Description

QwQ-32B is a 32-billion-parameter large language model designed to deliver high-level reasoning and intelligence. It achieves performance comparable to DeepSeek R1, a 671-billion-parameter model (with 37 billion activated), highlighting the efficiency of well-optimized foundation models trained on extensive world knowledge. The model incorporates agent-like capabilities, allowing it to perform critical reasoning, utilize tools, and adapt its behavior based on real-time environmental feedback. These features enable QwQ-32B to handle complex tasks with deep thinking and dynamic decision-making.

⁠Intended uses

QwQ-32B is designed for tasks requiring advanced reasoning and problem-solving abilities.

• Mathematical problem solving: Excels in complex mathematical computations and proofs.
• Code generation and debugging: Assists in writing and troubleshooting code across various programming languages.
• General problem-solving: Provides insightful solutions to diverse challenges requiring logical reasoning.

⁠Characteristics

⁠Available model variants

¹: VRAM estimated based on model characteristics.

⁠Use this AI model with Docker Model Runner

First, pull the model:

docker model pull ai/qwq

Then run the model:

docker model run ai/qwq

For more information on Docker Model Runner, explore the documentation⁠.

⁠Considerations

• Language mixing and code-switching: The model may unexpectedly switch languages, affecting response clarity.
• Recursive reasoning loops: Potential for circular reasoning patterns leading to lengthy, inconclusive responses. Use Temperature=0.6, TopP=0.95, MinP=0 to avoid this and use TopK between 20 and 40 to filter out rare token occurrences while maintaining the diversity of the generated output.
• Performance limitations: While excelling in math and coding, it may underperform in common sense reasoning and nuanced language understanding.

⁠Links

• QwQ-32B: Embracing the Power of Reinforcement Learning⁠