Meta Open Sources Code Llama to Challenge AI Coding Assistants
Meta releases Code Llama, an open source AI model designed to generate, explain, and debug code across multiple popular programming languages. The new system builds directly on the company's recently released Llama 2 framework.
Meta releases Code Llama, an open source AI model that generates and explains computer code in natural language. This new system competes directly with established tools like GitHub Copilot and Amazon CodeWhisperer by completing and debugging software across popular languages such as Python, C++, Java, and PHP.
The model is based on Meta's Llama 2 framework but focuses heavily on code-related training data to produce higher quality programming output. It comes in several variations, including a version specifically optimized for Python and another fine-tuned to understand complex natural language instructions from users.
Meta trains the Code Llama models using 500 billion tokens of code and provides options ranging from 7 billion to 34 billion parameters. The company argues that an open approach to AI coding models drives both innovation and safety by allowing the broader community to evaluate capabilities and fix vulnerabilities.