Summary: Code Llama is a code-specialized variant of Llama 2 fine-tuned/trained on code-focused datasets to improve code generation, completion, infilling, and code-aware NL responses. The model family includes multiple parameter sizes (7B, 13B, 34B) and several flavors (base, Python-fine-tuned, and instruction-tuned/Instruct-chat styles). Supported tasks and languages include code generation, completion, debugging assistance, and NL↔code prompts across many languages (Python, C/C++, Java, TypeScript/JavaScript, PHP, C#, Bash, etc.). Typical uses are local inference, code assistants and IDE integrations, research, and experimentation, subject to license terms. Sources and availability (note: Meta ai.meta.com pages were reported as temporarily down in the source text): recommended authoritative sources are the Hugging Face CodeLlama organization and individual Hugging Face model cards (these include downloads, example code, and safety/usage notes). Community mirrors and conversions (GGUF/GGML) and projects such as llama.cpp provide local/embedded runtime options. Meta/GitHub repos for the Llama family and consolidated tooling may also host relevant resources; when Meta’s official pages are available again, consult them for official downloads and license text. Licensing and safety: Use of Code Llama models is governed by Meta’s license. Because the official Meta license page was reported as temporarily unavailable, consult Hugging Face model cards and the license text on ai.meta.com once it is accessible. Hugging Face model cards and mirrors refer to a "Meta license" / "custom commercial license" — do not assume unrestricted use; verify the license before commercial deployment. Model cards also warn that outputs can be incorrect, biased, or unsafe; perform safety testing, appropriate fine-tuning/instruction tuning, and deploy guardrails for hallucinations, data leakage, and harmful content. Technical notes: Model cards report training/fine-tuning on code-focused data and large-scale compute; for precise training details, consult each model’s Hugging Face model card or official announcement. Hugging Face model cards also include example integration snippets (Transformers/tokenizer/model load and generation/pipeline usage) for typical local or cloud inference workflows. Immediate suggested next steps while the Meta site is down: (1) Retry the Meta URL later or check Meta AI status/official channels for official wording and license; (2) Use Hugging Face CodeLlama pages for downloads, model cards, example code, and safety guidance (links below); (3) For local deployment, examine community GGML/GGUF builds and llama.cpp; (4) Before production/commercial use, confirm license terms with the official Meta license page or legal counsel when ai.meta.com is available; (5) Optionally have someone monitor the Meta page and notify when it is restored to extract any unique official information.