Qwen2.5-7B-Instruct

Qwen2.5-7B-Instruct is a 7.61-billion-parameter instruction-tuned decoder-only language model from Alibaba's Qwen team, released on September 19, 2024 as part of the Qwen2.5 open-source family. Built on an expanded pre-training corpus (up to 18 trillion tokens) and refined with supervised fine-tuning and alignment, it delivers strong performance across general knowledge, mathematics, coding, and multilingual tasks spanning 29+ languages. It supports up to 131,072 tokens of context (128K) via YaRN extrapolation and can generate up to 8,192 tokens per pass.

Official benchmarks show Qwen2.5-7B-Instruct significantly outperforming peers in its size class, including Gemma2-9B-IT and Llama-3.1-8B-Instruct. It scores 75.4 on MMLU-redux, 91.6 on GSM8K, 84.8 on HumanEval, 75.5 on MATH, and 56.3 on MMLU-Pro. It also achieves an MT-Bench score of 8.75 and Arena-Hard score of 52.0, reflecting strong human-preference alignment. The model excels particularly in math and code generation relative to its parameter count.

Licensed under Apache 2.0, Qwen2.5-7B-Instruct is widely available on Hugging Face, Ollama, and vLLM, making it a popular choice for local deployment, fine-tuning, and production chat applications. Its efficient size-to-capability ratio has made it one of the most recommended open-weight 7B-class models in the community.