UPDATE: The model was pulled back after several hours due to some issues with the Stop function. It is expected that it will become available in upcoming days.
Gemma V2 27b model started appearing on Google AI Studio in a Preview mode. Gemma 2 is Google's next-gen open-weights model with 27B parameters, a major upgrade from the previous 2B and 7B versions. Despite its size, it matches 70B models like Llama while using half the compute.
Key Gemma V2 27B features:
- Increased performance on benchmarks like MMLU, HellaSwag, GSM8K, often outperforming 70B Llama 3 and 314B Grok-1.
- Versatile tuning across Google Cloud, Hugging Face, TensorRT for specific apps.
- Optimized for NVIDIA GPUs and Google TPUs for efficient, cost-effective deployment.
- Incorporates safety via automated filtering of sensitive data and RLHF fine-tuning.
Future developments:
- PaliGemma for vision-language tasks like image captioning.
- More Gemma variants for diverse apps.
- Fostering collaborative community.
Gemma V2 offers powerful capabilities with responsible practices. Google provides tools/resources to build, tune, deploy efficiently, accelerating AI innovation.