Meta Unveils Open-Source Code Llama to Accelerate Software Development
Meta releases Code Llama, a specialized large language model built on Llama 2 that generates and discusses code. The open-source tool supports multiple programming languages and is available in three sizes for various latency needs.
Meta releases Code Llama, a state-of-the-art large language model specifically designed to generate and discuss code. Built as a specialized version of Llama 2, this new model trains on an massive dataset of 500 billion code tokens to deliver enhanced coding capabilities. It supports popular programming languages like Python, C++, Java, and PHP, while functioning as both a productivity tool for seasoned developers and an educational resource for coding beginners.
The company offers Code Llama in three distinct sizes—7B, 13B, and 34B parameters—to address different latency and serving requirements. The largest 34B model provides the most advanced coding assistance, whereas the smaller 7B and 13B models operate on a single GPU for real-time, low-latency tasks like code completion. Additionally, the 7B and 13B models feature fill-in-the-middle capabilities that allow them to seamlessly insert code into existing projects.
Staying true to an open approach to artificial intelligence, Meta releases Code Llama for both research and commercial use under the same community license as Llama 2. Beyond the base models, the release includes specialized variations like Code Llama – Python, which receives additional fine-tuning on 100 billion tokens of Python code. These specialized models aim to provide extra utility for the AI community, where Python and PyTorch play a crucial role.