Meta has released the Code Llama model, which is a 70B model. This release follows their largest model, which was only 34B, released a year ago. They have not disclosed any further details, except for mentioning that their HumanEval score is 67.8, surpassing GPT-4 (67.0) and Gemini Pro (67.7), with other systems typically using a combination of GPT-4/GPT-3.5 and other techniques to achieve higher scores.
The released model consists of three sub-models: a basic model for general usage, a Python-specific model, and an Instruct model for programming instruction. The Instruct model is the one that reported the HumanEval score.
These models are available for free but come with limitations. They cannot be used by organizations with more than 700 million monthly active users.
TLDR: Meta has launched the Code Llama model, a 70B model with a HumanEval score of 67.8, outperforming GPT-4 and Gemini Pro. It consists of three sub-models and is available for free with usage restrictions.
Leave a Comment