Meta releases Code Llama: Llama 2 models for code

AI Impact Summary

Meta has released Code Llama, a family of Llama 2 models specifically trained for code-related tasks. These models, ranging from 7B to 34B parameters, demonstrate state-of-the-art performance across various programming languages and offer capabilities like code completion, infilling, and unit test generation. The models' unique context window of 16k (extrapolating to 100k) and training data, including a large code dataset and natural language discussions, represent a significant advancement for software engineers seeking productivity boosts.