"Code Llama: Open Foundation Models for Code." Meta's first major code-specialized open-weight family, built on Llama 2 with continued pretraining on a code-heavy corpus and long-context fine-tuning to 100K tokens. Available at 7B, 13B, 34B, and 70B sizes, with base, Python-specialized, and Instruct variants.

Foundational for the open-coding-model lineage. Adopted as the base for many downstream code models (DeepSeek-Coder ancestry, Phind-CodeLlama, the Ziya-Coding-34B fine-tune, etc.) and one of the first open models to credibly approach GPT-4 on HumanEval at the 34B scale.

Model Details

Architecture DENSE
Parameters 70B
Base model llama-2

Variants

Name Parameters Notes
Code Llama 7B 7B
Code Llama 13B 13B
Code Llama 34B 34B
Code Llama 70B 70B

Paper

open-weightcodingfoundational

Related