Google Gemma 2 Models: 9B and 27B Open Weights

In this video, Sam Witteveen introduces Google’s new Gemma 2 models, available in 9 billion and 27 billion parameter versions. He discusses their performance, setup, and how they compare to other models like Llama 3. The 9 billion parameter model outperforms Llama 3’s 8 billion model on several benchmarks and fits on smaller GPUs like L4 and T4. The 27 billion parameter model is competitive with models up to 70 billion parameters but requires high-end hardware like Nvidia H100 or A100 GPUs. Sam highlights the models’ commercial licenses, their training on TPU 4s and 5s, and the extensive training data used. He also explores the models’ architecture changes, including a larger tokenizer with 256,000 tokens. The video includes a walkthrough of the technical report, emphasizing the models’ benchmarks and performance in the LMC Chatbot Arena. Sam demonstrates the models’ outputs using a notebook, showing their capabilities in generating high-quality, detailed responses and creative writing. He concludes by mentioning other Google AI releases, such as Gemini 1.5 Pro with a 2 million context window and code execution features, available on AI Studio.