Meta AI has released Code Llama, a machine-learning system designed to generate and elucidate code using natural language, particularly English. This technology is built upon Meta’s Large Language Model called Llama 2. Code Llama operates under the same community license as Llama 2, making it accessible for research and commercial purposes.
Code Llama aims to help software developers by generating natural language responses to code and prompts. Code Llama is set to launch in three distinct sizes, boasting 7 billion, 13 billion, and 34 billion parameters, respectively.
The smaller 7 billion and 13 billion models are geared towards scenarios requiring low latency, such as real-time code completion. On the other hand, the 34 billion model delivers superior overall results, albeit demanding more computing power.
Meta has also made refinements to the Python and Instruct versions of Code Llama. The Python variant boasts enhanced capabilities for Python code generation tasks. At the same time, the Instruct version has been fine-tuned to provide safer and more user-friendly responses to natural language prompts.
Meta asserts that Code Llama’s performance surpasses publicly accessible LLMs, as indicated by benchmark assessments. However, the company does not specify the exact models utilized for comparison. Notably, Code Llama achieved a commendable 53.7% on the HumanEval code benchmark, demonstrating its ability to accurately generate code based on textual descriptions.
The domain of code generators has long facilitated developers’ tasks. GitHub’s introduction of Copilot X, powered by OpenAI’s GPT-4 in March, streamlines code creation and validation. Beyond writing and checking code, GitHub Copilot is also adept at rewriting and updating existing code. Amazon’s AWS boasts CodeWhisperer, which produces, verifies, and enhances code. While Google has introduced AlphaCode, a coding tool, its official release is still pending.
