Meta’s newest replace to its code technology AI mannequin, Code Llama 70B, is “the biggest and best-performing mannequin” but. Code Llama tools launched in August and are free for each analysis and industrial use. In response to a publish on Meta’s AI weblog, Code Llama 70B can deal with extra queries than earlier variations, which implies builders can feed it extra prompts whereas programming, and it may be extra correct.
Code Llama 70B scored 53 p.c in accuracy on the HumanEval benchmark, performing higher than GPT-3.5’s 48.1 p.c and nearer to the 67 p.c mark an OpenAI paper (PDF) reported for GPT-4.
Constructed on Llama 2, Code Llama helps builders create strings of code from prompts and debug human-written work. Meta concurrently launched two different Code Llama instruments final fall, Code Llama – Python and Code Llama – Instruct, which centered on particular coding languages.
Code Llama 70B is out there on three variations of the code generator and continues to be free for analysis and industrial makes use of. The massive mannequin was skilled on 1TB of code and code-related knowledge. It’s hosted on the code repository Hugging Face, which affords entry to GPUs to run AI fashions.
Meta mentioned its bigger fashions, 34B and 70B, “return the very best outcomes and permit for higher coding help.”