Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), in collaboration with Petuum and LLM360, has unveiled "K2-65B," a revolutionary 65-billion parameter open-source large language model (LLM). K2-65B sets new benchmarks for transparency and performance in open-source AI, offering a comprehensive blueprint for documenting and examining the entire lifecycle of LLMs, including reproduction details.
Leveraging LLM360’s framework, this model promotes community-driven efforts toward artificial general intelligence (AGI) through transparent, peer-reviewed, reproducible, and collaborative open-source research and development. Available globally under the Apache 2.0 license, K2-65B is the first reproducible LLM to surpass leading private sector models like Llama 2 70B.
Large language models are pivotal in natural language processing (NLP), enabling computers to interpret and generate text akin to human communication. UAE companies have made significant strides in this field, notably with the creation of Jais, the world’s most advanced Arabic LLM, developed by Core42, MBZUAI, and Cerebras Systems.
Eric Xing, President of MBZUAI, emphasized that the launch of K2-65B reflects the UAE's advancing expertise in LLM development. He highlighted the importance of an open, collaborative approach in creating high-performing, efficient LLMs that can transform various sectors.
K2-65B underwent a rigorous two-stage training process and was evaluated through 22 multidisciplinary assessments, excelling in diverse domains such as math, coding, and medicine, outperforming Llama 2 70B in each area. On the Open LLM Leaderboard, K2-65B showcased superior performance metrics. Additionally, the chat version, K2-Chat, outperformed Llama 2 70B Chat in all evaluated aspects, demonstrating its capability to understand and generate human-like responses in various scenarios.
