Alibaba Cloud QianWen Unveils Another重量级 Open Source Model: Qwen1.5-32B with 32 Billion Parameters
Latest news shows that Alibaba Cloud's subsidiary, Tongyi Qianwen, has announced a major move to open source its 32 billion parameter model, Qwen1.5-32B. This model achieves a remarkable balance in terms of performance, efficiency, and memory usage, providing enterprises and developers with a more cost-effective model choice. It is worth noting that Tongyi Qianwen has already open sourced 7 large language models, accumulating over 3 million downloads worldwide, which fully demonstrates its wide-ranging influence and popularity.
Prior to this, Tongyi Qianwen has successfully open sourced 6 large language models with parameters ranging from 500 million to 720 billion, all of which have been upgraded to version 1.5. The smaller models have been widely used in edge deployment due to their convenience, while the 720 billion parameter model has repeatedly topped authoritative model rankings such as HuggingFace.
The newly open sourced 32 billion parameter model, Qwen1.5-32B, further achieves an ideal balance between performance, efficiency, and memory usage. Compared to the 14 billion model, the 32 billion model demonstrates more powerful capabilities in intelligent scenarios. Compared to the 72 billion model, the 32 billion model has lower inference costs, providing more cost-effective solutions for enterprises and developers. The Tongyi Qianwen team hopes that this 32 billion open source model can provide better solutions for downstream applications.
In terms of basic capabilities, the Tongyi Qianwen 32 billion parameter model has achieved outstanding results in multiple evaluations such as MMLU, GSM8K, HumanEval, and BBH. Its performance is comparable to the 720 billion parameter model and far exceeds other 300 billion level parameter models. This achievement once again proves the strong capabilities of Tongyi Qianwen in the field of large language models.
In terms of Chat models, the Qwen1.5-32B-Chat model scored over 8 points in the MT-Bench evaluation, with a relatively small gap compared to Qwen1.5-72B-Chat, further highlighting its excellent capabilities in the dialogue domain.
In addition, in terms of multilingual capabilities, the Tongyi Qianwen team selected 12 languages including Arabic, Spanish, French, Japanese, and Korean for evaluation. The multilingual capability of Qwen1.5-32B is only slightly inferior to the 720 billion parameter model, fully demonstrating its powerful cross-language processing ability.
By open sourcing the 32 billion parameter model, Qwen1.5-32B, Tongyi Qianwen not only enriches its open source model library but also provides more choices for enterprises and developers. At the same time, this also indicates that Tongyi Qianwen's technical strength in the field of large language models has been further enhanced, laying a solid foundation for future development.