one year on
Meta releases Code Llama, a code-generating AI model built on Llama 2
The new open-weight model family is specialized for programming tasks and free for commercial use, positioning Meta to compete with GitHub Copilot and other AI coding assistants.
Meta today released Code Llama, a family of large language models specialized for code generation, completion, and debugging. Built on top of Llama 2, Code Llama comes in three variants: a foundational code model, a Python-specialized version, and an instruction-tuned model designed to follow natural language prompts. The models range from 7 billion to 34 billion parameters and are trained on 500 billion tokens of code and code-related data, with the Python version receiving an additional 100 billion tokens of Python code. Code Llama supports up to 100,000 tokens of context, allowing it to handle large codebases. Meta claims the 34B model outperforms all other open-source code generators on the HumanEval and MBPP benchmarks, scoring 53.7% and 56.2% respectively, on par with ChatGPT.
All Code Llama models are released under the same community license as Llama 2, free for research and commercial use. Developers deploying the model on platforms with over 700 million monthly active users must request a separate license. Meta frames the release as a boost for open innovation in AI coding tools, but acknowledges risks: the models can generate malicious code if given deceptive prompts, and the company has only conducted limited internal red-teaming with 25 employees. The 7B model can run on a single GPU, while larger variants require more powerful hardware.
Code Llama arrives amid a surge of AI coding assistants. GitHub reports over 400 organizations using Copilot, and a Stack Overflow survey found 70% of developers already using or planning to use AI coding tools this year. Code Llama lands in a crowded field of AI coding tools, including GitHub Copilot and Amazon CodeWhisperer, while Meta says developers should perform safety testing and tuning before deployment.
The record
Reports that Code Llama, though internally red-teamed by 25 employees, can still generate harmful code when prompts are phrased benignly, e.g. requesting a script to encrypt all files in a home directory.
One year later — open only if you can handle spoilers
Code Llama became a foundational open-weight model for code generation, spawning numerous fine-tuned variants and research projects. However, its safety limitations were replicated in downstream models, and the community continued to debate the trade-offs of open versus closed AI coding tools. The 70B version arrived in January 2024, further solidifying its place in the ecosystem.