Qwen: Qwen3 8B (retires Oct 9)

Qwen3 8B is a dense causal language model built with 8.2 billion parameters, designed to serve as a flexible solution for both complex analytical tasks and everyday conversation. Its architecture features 36 layers and utilizes Grouped Query Attention to balance performance and efficiency. A defining characteristic of this model is its ability to toggle between a specialized thinking mode—optimized for mathematics, coding, and logical inference—and a non-thinking mode for standard dialogue. This dual-mode design allows users to adapt the model's behavior to the specific demands of a task, ensuring high-quality outputs whether the goal is creative writing, role-playing, or solving intricate agent-based problems.

Developed through extensive pre-training and post-training, the model demonstrates significant advancements in instruction following and human preference alignment. It is engineered to excel in multilingual environments, providing robust support for over 100 languages and dialects. Beyond its core conversational strengths, the model is built for practical agent integration, allowing it to interact effectively with external tools. With a native context window of 32,768 tokens that can be extended to 131,072 tokens using YaRN scaling, the model is well-positioned for long-context applications, making it a powerful tool for developers and users seeking a balance of reasoning depth and operational versatility.