Meta releases Code Llama to take on coding tasks

Code Llama is an open-source large language model designed specifically to write, complete, and debug code. You can prompt it in plain English to build functions from scratch or feed it existing files to track down bugs and explain logic.

Meta built the model by training Llama 2 on an extra 500 billion tokens of code and technical data. It supports most popular languages right out of the box, including Python, C++, Java, PHP, TypeScript, C#, and Bash.

What it can do

The model comes in three flavors: the base Code Llama, a Python-specialized edition, and an instruction-tuned variant called Code Llama - Instruct. If you plan to talk to the model using natural prompts, Meta explicitly recommends sticking to the Instruct version.

Context size is one of the standout upgrades. While base Llama 2 caps out at 4,096 tokens, Code Llama can handle up to 100,000 tokens of context. That lets you drop an entire codebase or large log files into the prompt for deep debugging.

For smaller sizes, Meta also baked in fill-in-the-middle capabilities. This allows the 7B and 13B models to insert code between two existing blocks rather than just generating new code from top to bottom, making them ideal for IDE autocomplete features.

Specs and hardware

Code Llama is available in four parameter sizes: 7B, 13B, 34B, and a 70B edition trained on 1 trillion tokens.

Hardware requirements scale cleanly with the sizes. The 7B model fits on a single consumer GPU, making it fast enough for real-time code completion with minimal latency.

The 34B and 70B variants trade raw speed for reasoning capability. On the HumanEval benchmark, Code Llama 34B hit 53.7%, and on MBPP it reached 56.2%, matching ChatGPT and easily topping existing open alternatives.

All models share the permissive Llama 2 community license, meaning they are free to download and deploy for research and commercial software.

Why it matters

Code Llama gives developers a serious alternative to closed services like GitHub Copilot or OpenAI API calls. If you work in an environment where proprietary code cannot leave the local network, you can now run a capable coding assistant entirely offline.

By matching commercial-grade coding accuracy on an open-weight license, Meta puts heavy pressure on paid developer tools. We expect to see an explosion of local IDE plugins, testing bots, and self-hosted programming assistants built directly on top of these weights.