Meta has released Code Llama, a new AI tool built on top of its large language model Llama 2. Code Llama is designed to help developers by generating new code and debugging existing code. The model is available under an open community license, allowing free use for research and commercial purposes.
Capabilities of Code Llama
According to Meta, Code Llama can produce code strings based on text prompts. It can also autocomplete code when given a specific code snippet to work with.
In addition to the base model, Meta released a Python-specialized version called Code Llama-Python. It also released an instruction-focused model called Code Llama-Instrct that understands natural language instructions. Meta notes each version of the model is tailored to specific tasks. For example, the base and Python models do not understand natural language well.
“Programmers are already using LLMs to assist in a variety of tasks, ranging from writing new software to debugging existing code,” Meta stated in a blog post. “The goal is to make developer workflows more efficient so they can focus on the most human-centric aspects of their jobs.”
Benchmark Testing Performance
According to Meta, Code Llama outperformed other publicly available LLMs in benchmark testing. However, the company did not name the specific models tested against. It said the model achieved 53.7% on the HumanEval code benchmark. It also claimed the tool could accurately write code based on text descriptions of desired functionality. Meta will release three sizes of Code Llama and said its smallest size fits on a single GPU for more low-latency projects.
The Rise of AI Coding Assistants
Code generators promise huge gains but aren’t risk-free. These tools offer clear appeal – programmers and non-programmers alike stand to benefit. GitHub says over 400 organizations now use Copilot, with developers coding 55% faster than before. A Stack Overflow survey found 70% already employ AI coding tools or plan to this year, citing productivity and learning boosts.
Code generation tools like GitHub’s Copilot, AWS CodeWhisperer, and Google’s AlphaCode aim to expedite coding and maintenance tasks using AI. However, legal concerns remain about potential copyright infringement by these tools. GitHub and its parent Microsoft were recently sued for allegedly reproducing licensed code through Copilot.
The release of Code Llama reflects Meta’s push into this emerging market for AI developer assistants. With its open license, it could see rapid adoption. However, Meta will also need to ensure it avoids legal pitfalls in how the model derives and generates code.