Code Llama

Code Llama is Meta AI's family of downloadable coding language models derived from Llama 2. It is a model resource for developers building their own applications, rather than a hosted chat subscription. The official announcement and model card distinguish base models for code synthesis, Python-specialized variants and Instruct variants for instruction-following assistance. Inputs and outputs are text; described uses include code completion, explaining code and debugging assistance.

The official code repository currently resolves to meta-llama/codellama and is marked archived and read-only since July 2025. Its public inference examples remain visible. An accessible repository does not establish continued maintenance, a working hosted endpoint or availability of every model download. This review read public documentation and license files without installing dependencies, requesting model weights or running inference.

Meta's announcement describes research and commercial availability, while the linked Llama 2 Community License grants limited royalty-free rights subject to conditions. Redistribution requires the agreement and attribution notice; certain large services require additional permission. The acceptable-use policy applies, and outputs may involve third-party licenses. Free licensing does not include GPU hardware, hosting, electricity or third-party inference-provider fees, and it should not be presented as an unrestricted permissive software license.

The model card describes a static model trained on an offline dataset and warns of inaccurate or objectionable outputs. English and programming-language evaluation does not establish reliable performance in every language or application. Context support and infilling differ between variants; consult the selected model's documentation rather than applying a family-wide headline limit. Developers should review generated code, check dependencies and licenses, and perform application-specific safety and correctness testing before deployment.