Meta Debuts ‘Code Llama’ Generative AI Model for Writing and Explaining Software
Meta has released a new large language model for generating, editing, and discussing code in multiple programming languages. Code Llama, as the name indicates, extends the Llama 2 LLM released last month by Meta into software development with an eye toward enhancing productivity and lowering barriers to entry for new programmers.
From the Llama 2 skeleton, Meta trained Code Llama on hundreds of billions of tokens of code from GitHub and other sources, allowing it to understand and generate code in languages like Python, Java, C++, and more. Meta employed the same data for training Llama 2. The difference comes from how the AI zeroed in on any of the code-related information for a narrower but deeper set of capabilities. The AI can complete half-written functions, explain code snippets in plain language, and debug errors. While Code Llama is the foundational model, Meta released two additional variants at the same time. Codel Llama is a Python-focused model, and Code Llama – Instruct is designed specifically for understanding natural language instructions.
Meta is releasing Code Llama in three model sizes – 7 billion, 13 billion, and 34 billion parameters – to suit different latency and capability requirements. The two smaller models have also been trained specifically for code completion, allowing them to readily fill in missing sections of code, while the largest 34 billion parameter model achieves the strongest performance overall. Meta claims Code Llama beats any other publicly available LLM when it comes to coding.
“Code Llama has the potential to be used as a productivity and educational tool to help programmers write more robust, well-documented software,” Meta explained in its announcement. “We believe an open approach to AI is best for developing new AI tools that are innovative, safe and responsible, so we’re releasing Code Llama for both research and commercial use under the same community license as Llama 2.”
Code Llama joins the growing ranks of AI coding assistants, which have become increasingly popular. GitHub Copilot has been at the forefront, but Google’s Codey, Amazon’s CodeWhisperer, and Hugging Face’s Starcoder all offer variations on the theme of a tool that can reduce the time and cost of software development by writing and explaining code.