Meta launches code Llama 70B to challenge GitHub Copilot

Researchers and developers may now download Code Llama, Meta's most powerful and free large language model as Meta's AI coding rival to GPT-4. This generative AI model has 70 billion parameters and was first revealed in August 2023.

"We are open-sourcing a new and upgraded Code Llama, which includes a bigger 70B parameter model. Writing and rewriting code has emerged as one of the most essential applications for AI models today. The ability to code has also been shown to be necessary for AI models to handle data in other fields more rigorously and rationally. In a Facebook post, Meta CEO Mark Zuckerberg stated, "I am proud of the progress here and look forward to including these advances in Llama 3 and future models as well."

According to Meta, Code Llama is the most advanced and best-performing model in the Llama family. It is available for free in three versions: CodeLlama – 70B, the foundational code model, CodeLlama – 70B – Python, a Python-specific version, and CodeLlama – 70B – Instruct 70B, a version fine-tuned with human instruction and self-instruction code synthesis.

Meta's most recent version of their code creation AI model, Code Llama 70B, is "the largest and best-performing model" ever. Code Llama tools were released in August and are free for study and commercial use. According to a post on Meta's AI blog, Code Llama 70B can answer more inquiries than earlier versions, allowing developers to provide more hints when programming and improving accuracy.

On the HumanEval benchmark, Code Llama 70B earned 53% accuracy, surpassing GPT-3.5's 48.1 % and approaching the 67% figure stated in an OpenAI study (PDF) for GPT-4.

Built on Llama 2, Code Llama enables developers to generate code strings from prompts and troubleshoot human-written work. Last autumn, Meta released two further Code Llama tools: Code Llama – Python and Code Llama – Instruct, which focused on specific coding languages.

Code Llama 70B is available in three versions of the code generator and is still free for study and commercial usage. The huge model was trained using 1TB of code and code-related data. It is hosted on Hugging Face, a code repository that provides access to GPUs for AI models.

Meta stated that its bigger models, 34B and 70B, "return the best results and allow for better coding assistance."

Other AI developers developed code generators within the last year. Amazon's CodeWhisperer debuted in April, while Microsoft used OpenAI's methodology to create GitHub Copilot.