On September 19, 2024, Alibaba’s Qwen team released Qwen2.5, one of the broadest open-weight model lineups to date. The base family spanned 0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B parameter sizes, accompanied by specialized Qwen2.5-Coder and Qwen2.5-Math variants. The team said the base models were pretrained “on our latest large-scale dataset, encompassing up to 18 trillion tokens,” with most weights released under the Apache 2.0 license.
Qwen2.5 supported context windows up to 128K tokens and generation up to 8K tokens, and covered more than 29 languages. The team reported that “compared to Qwen2, Qwen2.5 has acquired significantly more knowledge (MMLU: 85+)” along with notable gains in coding and mathematics, with Qwen2.5-Coder trained on 5.5 trillion tokens of code-related data.
Qwen2.5 mattered because it gave developers a single, permissively licensed family covering everything from on-device 0.5B models to a 72B near-frontier model, with strong code and math specialists. That breadth helped make Qwen one of the most widely adopted open-weight bases worldwide and a serious challenger to Western open models.