Kuaishou’s Self-Developed Large-Scale Model ‘KwaiYii’ Makes Its Debut

Recently, Kuaishou’s self-developed large language model ‘KwaiYii‘ has entered internal testing and provided standard APIs and customized project collaboration solutions for the business team.

In the CMMLU Chinese language-oriented foundation model list, the impressive 13B version KwaiYii-13B ranks first in both five-shot and zero-shot categories. It demonstrates strong performance in humanities, specific Chinese topics, and achieves an average score of over 61 points.

Upon searching the GitHub page, it was found that the official description states: KwaiYii, developed independently by the Kuaishou AI team, is a series of large-scale language models (LLM) built from scratch. Currently, it includes models with various parameter sizes, such as the pre-training model (KwaiYii-Base) and the chat model (KwaiYii-Chat). Here we introduce the KwaiYii-13B series model, which has a scale of 13 billion parameters.

Its main features include: KwaiYii-13B-Base pre-trained model has excellent general technical capabilities and achieves state-of-the-art performance in most authoritative Chinese/English benchmarks with the same model size. For example, the KwaiYii-13B-Base pre-trained model is currently leading in benchmarks such as MMLU, CMMLU, C-Eval, HumanEval at the same model scale.

SEE ALSO: Kuaishou Incentivizes User Collaboration with 60 Billion Network Traffic

KwaiYii-13B-Chat dialogue model has excellent language understanding and generation capabilities, supporting a wide range of tasks such as content creation, information consultation, mathematical logic, code writing, and multi-turn conversations. The results of manual evaluation show that KwaiYii-13B-Chat surpasses mainstream open-source models and approaches the level of ChatGPT (3.5) in content creation, information consultation, and mathematical problem-solving.

According to reports, the AI team at Kuaishou will continue to iterate on the ‘KwaiYii’ large model. On the one hand, they will continue to optimize model performance and develop multimodal capabilities. On the other hand, they are also promoting implementation in more C-end and B-end business scenarios.