Code Llama 70B: Meta releases 'largest', 'best-performing' AI coding model

Code Llama 70B achieved a 53 per cent accuracy on HumanEval benchmark
An image of Code Lama app. — Meta/File
An image of Code Lama app. — Meta/File

Meta has introduced Code Llama 70B, the latest update to its open-source artificial intelligence (AI) coding model described as "the largest and best-performing model in the Code Llama family," 

The AI model achieved a 53pc accuracy on the HumanEval benchmark, approaching the capabilities of OpenAI's GPT-4, which scored 67pc, according to the company’s report. Meta's coding models now include Code Llama 7B, Code Llama 13B, Code Llama 34B, and the new Code Llama 70B.

According to Gadgets360, Meta CEO Mark Zuckerberg announced the release in a Facebook post, stating, “We're open-sourcing a new and improved Code Llama, including a larger 70B parameter model. Writing and editing code has emerged as one of the most important uses of AI models today. [..] I'm proud of the progress here, and looking forward to including these advances in Llama 3 and future models as well.”.

Read more: Meta's Threads app downloads shoot up, surpass X

Code Llama 70B is available in three versions, the foundational model, Code Llama - Python (for the Python programming language), and Code Llama - Instruct (with natural language processing capabilities, it is usable without coding knowledge).

The AI coding assistant generates both code and natural language responses, facilitating code explanation and answering related queries. Trained on 1 trillion tokens of coding and code-related data, the 70B model is free for research and commercial purposes, hosted on the Hugging Face coding repository.

Benchmark scores indicate the model's accuracy compared to rival coding-focused AI models. On the HumanEval benchmark, Code Llama 70B scored 53pc, surpassing OpenAI's GPT-3.5 (48.1pc). 

On the Mostly Basic Python Programming (MBPP) benchmark, it achieved 62.4pc, outscoring GPT-3.5 (52.2pc). Although GPT-4 only released its HumanEval accuracy scores online, it scored 67pc, surpassing Code Llama 70B.

Meta previously released Code Llama in August 2023, based on the Llama 2 foundational model. It accepts both code and natural language prompts, generating responses in both forms. Code Llama facilitates code generation, editing, analysis, and debugging, with its Instruct version aiding users in understanding codes through natural language.