Alibaba Cloud, the digital technology and intelligence backbone of Alibaba Group, has announced its latest contribution to the open-source community by open-sourcing its 7-billion-parameter Large Language Models (LLM), Qwen-7B and Qwen-7B-Chat, through its AI model community ModelScope, and the collaborative AI platform Hugging Face.

Alibaba Cloud introduced its proprietary LLM, Tongyi Qianwen, earlier this year in April. This cutting-edge model, capable of generating human-like content in both Chinese and English, has different model sizes, including seven billion and above parameters. This time, the open-source release includes the pre-trained 7-billion-parameter model, Qwen-7B, and its conversationally fine-tuned version, Qwen-7B-Chat.

In an effort to democratize AI technologies, the models’ code, model weights, and documentation will be freely accessible to academics, researchers and commercial institutions worldwide. For commercial uses, the models will be free to use for companies with fewer than 100 million monthly active users. Programs with more users can request a license from Alibaba Cloud.

“By open-sourcing our proprietary large language models, we aim to promote inclusive technologies and enable more developers and SMEs to reap the benefits of generative AI,” said Jingren Zhou, CTO of Alibaba Cloud Intelligence. “As a determined long-term champion of open-source initiatives, we hope that this open approach can also bring collective wisdom to further help open-source communities thrive.”

The Qwen-7B was pre-trained on over 2 trillion tokens, including Chinese, English and other multilingual materials, code, and mathematics, covering general and professional fields. Its context length reaches 8K. In training, the Qwen-7B-Chat model was aligned with human instructions. Both Qwen-7B and Qwen-7B-Chat models can be deployed on cloud and on-premises infrastructures. This enables users to fine-tune the models and build their own high-quality generative models effectively and cost-efficiently. 

The pre-trained Qwen-7B model distinguished itself in the Massive Multi-task Language Understanding (MMLU) benchmark, scoring a notable 56.7, outperforming other major pre-trained open-source models with similar scales or even some larger-size models. This benchmark assesses a text model’s multitask accuracy across 57 varied tasks, encompassing fields such as elementary mathematics, computer science and law. Moreover, Qwen-7B achieved the highest score among models with equivalent parameters in the leaderboard of C-Eval, a comprehensive Chinese evaluation suite for foundational models. It covers 52 subjects in four major specialities including humanities, social sciences, STEM and others. Additionally, Qwen-7B reached outstanding performance on benchmarks of mathematics and code generation, such as GSM8K and HumanEval.