Meta Unveils Open-Source Code Llama AI Model for Developers
Meta releases Code Llama, a state-of-the-art large language model designed specifically to generate, discuss, and debug code across multiple programming languages.
Meta releases Code Llama, a state-of-the-art large language model built specifically to generate and discuss code. Built upon Llama 2, this specialized tool accelerates developer workflows and lowers the barrier to entry for coding beginners. In keeping with their commitment to open AI, Meta makes Code Llama available for both research and commercial use under a community license.
The model supports popular programming languages like Python, C++, Java, and TypeScript, and handles tasks such as code completion and debugging. Meta offers three model sizes—7B, 13B, and 34B parameters—each trained on 500 billion tokens of code data. The 7B and 13B models include fill-in-the-middle capabilities for immediate code completion tasks, while the larger 34B model delivers the highest quality assistance for complex projects.
Beyond the base models, Meta introduces specialized variations including Code Llama – Python and Code Llama – Instruct. The Python variation fine-tunes the system further on 100 billion tokens of Python code to serve the AI community's heavy reliance on the language. These different model sizes allow developers to choose the best fit for their specific latency and serving requirements, whether they need real-time suggestions or deep coding support.