Meta’s New Code Llama 70B Rivals GitHub Copilot for Generative AI Programming Models
Meta has released a new, more powerful version of its Code Lllama model for writing and designing software that could be real competition for GitHub Copilot and other AI pair programmers. Code Llama 70B is a beefed-up version of the Code Llama large language model (LLM) introduced last August and born out of the Llama 2 LLM.
Code Llama 70B
Meta created Code Llama as a tool for enhancing productivity and lowering barriers to entry for new programmers. Code Llama 70B upscale the model to 70 billion parameters. That’s nearly twice the previously largest Code Llama model, which initially launched with 7 billion, 13 billion, and 34 billion- parameter options. The bigger size allows the model to handle more queries and contextual information than prior versions when assisting developers in writing and debugging code. Like its smaller siblings, Code Llama 70B can complete half-written functions, explain code snippets in plain language, and debug errors. The upgraded model was trained on over one terabyte of code data and is hosted on the code repository Hugging Face, which provides compute resources to run AI systems.
Two variations focusing on Python and instruction-based training were also released alongside the base 70B model. Meta emphasizes its larger 34 billion and 70 billion parameter Code Llama models, which offer the most advanced coding assistance capabilities. However, all model sizes are available for free public experimentation and application deployment. Meta claims the new model is both the largest and most accurate publicly available code-generating AI system. The tech giant described how Code Llama 70B achieved state-of-the-art results on standard programming benchmarks, including a 53% accuracy in code completion on the HumanEval test – surpassing Microsoft and OpenAI’s GitHub Copilot tool and nearing scores reported for GPT-4.
Meta claims that openly sharing its research encourages transparency and helps shape responsible AI development across the ecosystem. The company says it undertook extensive safety reviews before releasing Code Llama publicly. However, optimizing the societal impacts of code-generating systems remains an evolving challenge. Continued version upgrades suggest coding AI still requires significant improvement before matching human abilities. For now, Meta frames tools like Code Llama as assistants that can automate rote coding tasks and allow programmers to focus on higher-level thinking. The system outputs still generally require human review, testing, and oversight.
Meta’s interest in continually improving Code Llama and increasing its power fits with the rapid rise in demand for similar tools. While GitHub Copilot has arguably been the most successful, it has plenty of rivals beyond Meta, including Google’s Codey, Amazon’s CodeWhisperer, and Hugging Face’s Starcoder.