On August 24, 2023, Meta released Code Llama, a large language model designed specifically for programming tasks, under an open-source license that permits commercial use. The model is available in three parameter sizes—7 billion, 13 billion, and 34 billion—allowing developers to choose based on their hardware capabilities. The smallest version can run on a laptop, while the largest requires server-grade infrastructure.
Code Llama is built on Llama 2, Meta’s previous model generation, and represents a continuation of the company’s shifting approach to AI distribution. The original Llama model, released in February 2023, was restricted to academic research and non-commercial use only. After its weights leaked via BitTorrent, Meta loosened restrictions with Llama 2, releasing both foundation and instruction-tuned versions to a broader audience. Code Llama follows that trajectory.
The model can generate code from plain English prompts, debug existing code, and explain code functionality in natural language. Meta is not charging for Code Llama. The company has instead chosen to give it away under a license that allows commercial use—a far more permissive approach than the original Llama model’s license.
Timing plays a strategic role in this release. Software developers represent a key market for AI assistance tools. GitHub Copilot, powered by OpenAI’s models, charges a monthly subscription fee. Amazon CodeWhisperer offers a free tier but integrates closely with Amazon Web Services. Code Llama enters this landscape as a free, specialized alternative with no vendor lock-in beyond the model itself.
The three parameter sizes allow developers to balance speed against accuracy. The 7-billion-parameter model is fast and lightweight, while the 34-billion-parameter version is slower but more accurate. This design choice lets developers integrate Code Llama into existing workflows without major infrastructure changes.
Meta’s broader strategy involves betting that open models will win over the long term. By releasing Code Llama for free, the company undercuts competitors that charge per seat or per token. It also gains real-world usage data, bug reports, and community improvements without direct R&D investment. The commercial-use license allows companies to fine-tune Code Llama on proprietary codebases and build products on top of it, creating dependency on Meta’s ecosystem.
There are risks to this approach. Open models can be misused; bad actors could fine-tune Code Llama to generate malicious code. Meta is betting that the benefits of widespread adoption outweigh potential abuse. The company learned from the Llama leak that tight control invites circumvention. Now it is leaving the door open.
Code Llama represents an evolution of Meta’s strategy rather than a revolution. General-purpose LLMs can write code but often hallucinate syntax, lose context, or ignore precise instructions. Code Llama was built for one job, making it more specialized than multipurpose models. Whether that specialization is enough to displace established tools like Copilot remains an open question.
