Qwen2.5 7B Instruct

Qwen2.5 7B Instruct brings the September 2024 Qwen2.5 improvements into a 7B-parameter dense model that still supports function calling, structured JSON output, and tool use. Trained on roughly 18 trillion tokens with added math and code instruction, it reports 84.8 on HumanEval and 75.5 on MATH, outperforming peers like Gemma2-9B and Llama3.1-8B on core reasoning benchmarks. A 32K context window supports long-document processing and extended conversations, and multilingual coverage spans more than 29 languages. Its Apache 2.0 weights allow unrestricted fine-tuning and integration, making it a good fit for resource-constrained deployments, cost-sensitive production systems, and custom fine-tuning where the 72B size exceeds available capacity.