Overview

• The 270-million-parameter model combines 170 million embedding and 100 million transformer parameters to support strong instruction-following in a compact design.
• Quantization-Aware Training checkpoints enable INT4 deployment, and internal tests show 25 chat turns consumed just 0.75% of a Pixel 9 Pro battery.
• Google published pretrained, instruction-tuned, and QAT checkpoints and made them available through community hubs like Hugging Face, Ollama and services like Vertex AI to support rapid fine-tuning and deployment.
• Gemma 3 270M is distributed under the Gemma Terms of Use, granting broad commercial rights with restrictions but not qualifying as traditional open-source.
• Competitors have called for independent benchmarking in response to notable omissions in Google’s published IFEval comparisons.