Meta's new Code Llama AI programming tool raises the bar, outperforming GPT-3.5

Meta has released an upgraded version of its Code Llama AI programming tool, the Code Llama 70B, marking a significant leap in performance.

Introduced in August, the tool is now touted as Meta's "largest and best-performing model" to date. Code Llama 70B excels in generating and debugging extensive programming strings, surpassing its predecessors and even closing the gap with GPT-4.

The recent update showcases improved capabilities, enabling developers to input more prompts during programming, leading to enhanced accuracy. Code Llama 70B achieved a 53 percent accuracy score on the HumanEval benchmark, outperforming GPT-3.5's 48.1 percent.

Based on the Llama 2 architecture, Code Llama assists developers in generating and debugging code strings. The tool is part of Meta's broader Code Llama suite, which also includes specialized versions for Python and instructional coding. This latest iteration, Code Llama 70B, stands out as a free resource for both research and commercial use.