The next generation of LLaMa 2, a new open-source large language model by Meta, is now available for research and commercial use. This release boasts model weights and starting code for pre-trained and fine-tuned LLaMa language models, ranging from 7B to 70B parameters, catering to diverse applications.
With training on an extensive 2 trillion tokens, LLaMa 2’s pre-trained models feature double the context length compared to its predecessor, LLaMa 1. The fine-tuned models have been refined through over 1 million human annotations, enhancing performance across various external benchmarks, including reasoning, coding, proficiency, and knowledge tests.
Notably, LLaMa 2’s training was achieved using publicly available online data sources, and its fine-tuned model, LLaMa-2-chat, harnesses publicly available instruction datasets. The release has already received widespread support, with companies providing early feedback and expressing excitement to build with LLaMa 2 and researchers committed to advancing AI research with the platform’s capabilities. It is also stated that partnering cloud providers will incorporate the model into their offerings.