With GPT-4, Meta’s free Code Llama AI programming tool narrows the difference.


    Meta’s latest update to their code creation AI model, Code Llama 70B, is “the largest and best-performing model” ever. The August launch of Code Llama tools makes them available for free for both business and research use. Code Llama 70B can handle more queries than earlier iterations, which implies developers may give it more programming instructions and it can be more accurate, according to a post on Meta’s AI blog.

    On the HumanEval benchmark, Code Llama 70B achieved an accuracy score of 53 percent, outperforming GPT-3.5’s 48.1 percent and coming in closer to the 67 percent figure that an OpenAI article (PDF) for GPT-4 reported.

    Based on Llama 2, Code Llama assists programmers in debugging human-written work and generating code from prompts. Last autumn, Meta concurrently released two more Code Llama programs, Code Llama – Python and Code Llama – Instruct, which catered to particular programming languages.

    Three code generating versions support Code Llama 70B, which is still free for academic and business purposes. One terabyte of code and code-related data were used to train the huge model. It is housed in the Hugging Face code repository, which provides GPU access for AI model execution.

    The 34B and 70B, Meta’s largest models, “return the best results and allow for better coding assistance,” according to Meta.
    Within the last year, code generators were provided by other AI developers. Microsoft used OpenAI’s methodology to create GitHub Copilot, while Amazon released CodeWhisperer in April.