Qwen3-8B

Qwen3-8B is a dense 8.2-billion-parameter causal language model released by Alibaba's Qwen team on April 29, 2025 under the Apache 2.0 license. It is part of the Qwen3 family, which introduces hybrid thinking and non-thinking modes within a single model, allowing users to switch between deep chain-of-thought reasoning for complex math, coding, and logic tasks and fast direct responses for general chat. The model uses 36 layers with grouped-query attention (32 query heads, 8 KV heads) and supports 128K context natively, extendable to 131K tokens via YaRN rope scaling.

In the Qwen3 Technical Report (arXiv:2505.09388), Qwen3-8B in default thinking mode achieves 87.5 on MMLU-Redux, 97.4 on MATH-500, 76.0 on AIME'24, 57.5 on LiveCodeBench v5, and 85.8 on Arena-Hard, outperforming much larger models including Qwen2.5-14B-Instruct and DeepSeek-R1-Distill-Qwen-14B on most reasoning and alignment benchmarks. The base model scores 76.89 on MMLU, 89.84 on GSM8K, and 67.65 on EvalPlus (which includes HumanEval) in pre-training evaluations.

Qwen3-8B supports 119+ languages, strong agent and tool-calling capabilities via Qwen-Agent and MCP integration, and is widely deployed through Hugging Face, ModelScope, vLLM, SGLang, Ollama, and LM Studio. It has become one of the most downloaded open-weight models on Hugging Face, valued for delivering near-14B-class performance at 8B parameter scale with flexible reasoning control.