The original title is "Google's Gemma 2 is here. It's a big deal for open models."

Short summary

Google released Gemma 2 in 9B and 27B parameter sizes, with the 27B model offering performance competitive with models twice its size while running on a single GPU. The redesigned architecture uses interleaved local and global attention for better memory efficiency. Models are available on Hugging Face, Kaggle, and Google AI Studio with code examples for loading the instruction-tuned variants.

• •Gemma 2 launched in 9B and 27B sizes with redesigned hybrid attention architecture
• •27B model runs on a single H100 or A100 80GB GPU, making self-hosting more viable
• •Models available on Hugging Face, Kaggle, and Google AI Studio with PyTorch, JAX, and TensorFlow support

Generated with AI, which can make mistakes.