Meta Boosts Code Generation with New Code Llama Model

Facebook-parent Meta has published an improved version of its code generation model, Code Llama.

The latest version stands at 70 billion parameters in size, the largest thus far with prior ones at seven, 13 and 34 billion parameters. The new Code Llama comes in three versions – a base version, one that is fine-tuned for Python coding and a second instruct-tuned version tailored for natural language inputs.

Meta described the new Code Llama as “one of the highest performing open models available today.”

The models are open source and available under the same license as the previous versions, meaning it can be used for commercial applications – but to access them, you will need to fill out this form.

Meta CEO Mark Zuckerberg said he was “proud of the progress” in the new Code Llama and that its advances will be added to the upcoming Llama 3 and as well as future Meta models.

The original version of Code Llama dropped last August. Meta said this latest version is “the most performant base for fine-tuning code generation models” with a hope that the AI community build atop the new models.

The new versions blow the original version out of the water in terms of performance. The models also scored higher than rival code generation systems such as StarCoder, Codex from OpenAI, and Palm-Coder.

Related:With Code Llama, Meta Joins the Code Generation Party

The Instruct version even managed to achieve a score of 67.8 on the HumanEval Pass@1 test, which is 0.8 higher than GPT-4’s reported score.

Credit: Meta