Code Generation · Open Models · Language Models
Code Llama: Open Code Models Built from Llama
Code Llama adapts Llama-family models for code generation, infilling, and instruction-following, giving the open ecosystem stronger coding baselines.
Code Llama adapts Llama-family models for code generation, infilling, and instruction-following, giving the open ecosystem stronger coding baselines.
What problem it solves
Strong coding models were often closed or difficult to adapt. Code Llama addresses the need for open code-focused foundation models that can complete code, follow coding instructions, and support research and product experimentation.
The core method
Meta starts from Llama-family language models and continues training on code-heavy data. The release includes variants for general code generation, Python specialization, and instruction following, plus support for infilling so the model can edit code around a missing span.
Key results
Code Llama improves substantially over general-purpose Llama models on coding benchmarks and provides competitive open baselines at multiple model sizes. Its practical value comes from availability: teams can inspect, fine-tune, deploy, and compare against it more easily than closed models.
Why it matters
Open code models changed the coding assistant ecosystem. They enabled local experimentation, specialized fine-tuning, academic evaluation, and product prototypes where sending code to a closed service is not acceptable.
Limits and open questions
Open availability does not remove safety, licensing, or quality concerns. Coding models can generate insecure code, misunderstand project context, or overfit benchmark patterns. Real use still depends on tests, review, dependency awareness, and secure deployment practices.
One line: Code Llama gave open-source coding assistants a stronger foundation.