Code Llama was Meta's code-focused fine-tune of its Llama 2 foundation models, released in variants at 7B, 13B, 34B, and 70B parameters, each trained on roughly 500 billion additional code tokens with extended context windows reaching up to 100,000 tokens. It covered code generation, infilling, and instruction-following for programming tasks, and for a while was one of the strongest openly available code models alongside StarCoder.
It mattered as much for being open and genuinely usable as for raw benchmark performance: a well-resourced lab releasing a competitive code model with permissive enough terms for real use gave open-source developers a serious alternative to closed APIs. That mattered more in 2023 and 2024 than it does now that the broader open-model ecosystem has moved well past it.
The repository was archived in July 2025 and is no longer maintained, superseded by newer generations of Llama models that have absorbed and surpassed what Code Llama offered. It's now mostly of historical interest rather than a model anyone would choose to deploy today.