Meta deployed Code Llama in August 2023, a family of large language models available for free under a permissive license for research and commercial use, designed to generate, complete, and debug code across languages including Python, C++, Java, and Bash. Developers and companies can download and run the models to assist with programming tasks. Meta also uses a related internal coding tool called Metamate, which draws on both its own Llama models and OpenAI's GPT-4, according to Fortune reporting from December 2024.

Code Llama was first deployed on August 24, 2023, with models at 7B, 13B, and 34B parameters; a 70B variant followed on January 29, 2024. Each model is built on Llama 2 and trained on 500B tokens of code and code-related data, except the 70B model, which is trained on 1T tokens. The models accept text or code prompts as input and produce generated code or natural-language explanations of code as output, supporting languages including Python, C++, Java, PHP, TypeScript, C#, and Bash, with a 100K-token context window. Code Llama is available for download free of charge under a permissive community license that allows research and commercial use, though the Open Source Initiative has stated this license does not meet the Open Source Definition. Separately, Meta's internal coding tool, Metamate (originally called Code Compose), uses both Llama models and OpenAI's GPT-4 to assist Meta employees with coding tasks, according to two sources cited by Fortune in December 2024.