Llama 3

Llama 3 comes in 8B and 70B sizes, trained on over 15 trillion tokens of curated public data with a significant emphasis on code and reasoning. It incorporates Grouped-Query Attention for faster inference and is fine-tuned with supervised learning, rejection sampling, and policy optimization to enhance instruction following and preference alignment. The model can be run locally using llama.cpp, achieving around 71 tokens per second on a MacBook Pro M4 with 64GB RAM, enabling offline project work for hours without internet.

Key strengths: completely open-source, commercially permissible for organizations under $100M annual revenue, and capable of private, air-gapped deployment. It excels at English-heavy coding, content generation, and tool-calling tasks, while also supporting 30 languages (with limited non-English tokens).

Limitations include high memory requirements—70B version needs roughly 48GB RAM—making it unsuitable for low-resource devices. Chinese and other non-English performance is relatively weak due to training data imbalance.