Alibaba Cloud's Tongyi Qianwen open-source 32 billion parameter model Qwen1.5-32B can balance performance, efficiency, and memory usage to the greatest extent possible, providing enterprises and developers with a more cost-effective model choice. At present, Tongyi Qianwen has opened up 7 major language models, with a cumulative download volume exceeding 3 million in open source communities at home and abroad.
Tongyi Qianwen has previously opened up six large language models with parameters of 500 million, 1.8 billion, 4 billion, 7 billion, 14 billion, and 72 billion, all of which have been upgraded to version 1.5. Among them, several small-sized models can be easily deployed on the end side, while the 72 billion parameter model has industry-leading performance and has repeatedly appeared on models lists such as HuggingFace. This open-source 32 billion parameter model will achieve a more ideal balance between performance, efficiency, and memory usage. For example, compared to the 14B model, the 32B model has stronger capabilities in intelligent agent scenarios; Compared to 72B, 32B has lower inference costs. The Tongyi Qianwen team hopes that the 32B open source model can provide better solutions for downstream applications.
In terms of basic capabilities, the 32 billion parameter model of Tongyi Qianwen has performed well in multiple evaluations such as MMLU, GSM8K, HumanEval, BBH, etc. Its performance is close to that of Tongyi Qianwen's 72 billion parameter model, far exceeding other 30 billion parameter models.
In terms of the Chat model, the Qwen1.5-32B-Chat model scored over 8 points in the MT Bench evaluation, and the gap between it and Qwen1.5-72B-Chat is relatively small.
In terms of multilingual abilities, the Tongyi Qianwen team selected 12 languages including Arabic, Spanish, French, Japanese, Korean, etc., and conducted assessments in multiple fields such as exams, comprehension, mathematics, and translation. The multilingual ability of Qwen1.5-32B is only slightly inferior to the 72 billion parameter model of Tongyi Qianwen.