Code Llama 70B

Code Llama 70B was released by Meta on January 29, 2024 — announced personally by Mark Zuckerberg — as the largest member of the Code Llama family. Built on Llama 2 and further trained on 500B tokens of code plus long-context data, it shipped in three variants: the foundation model, a Python-specialized fine-tune, and an Instruct version for natural-language prompting. At release it scored 53% on HumanEval, beating GPT-3.5's 48.1% and approaching GPT-4's 67%, making it the strongest openly available code model of its moment.

The honest picture: it is a Llama 2-era model, and the code-model frontier has moved several generations past it — Qwen2.5-Coder, DeepSeek-Coder-V2, and newer models beat it soundly. The Llama community license is permissive for most commercial use but is not OSI open source. Still, it was a landmark release that proved open code models could compete with closed ones, and its Instruct variant remains a usable self-hosted coding assistant.