Google's Gemma provides lightweight open models with base and instruction-tuned versions in 2B and 7B sizes.