Now Available: Google Gemma 3n and Gemma 3 270M — On-Device AI Goes Multimodal

Two new models from Google DeepMind's Gemma family are live on the Hanzo AI Gateway.

Gemma 3n — Multimodal AI for Phones and Edge

Released February 11 as a developer preview, Gemma 3n is Google's first model designed natively for on-device deployment. It uses a MatFormer architecture with Per-Layer Embeddings to achieve remarkably small memory footprints:

• E2B — 5B total parameters, 2B effective, fits in 2 GB RAM
• E4B — 8B total parameters, 4B effective, fits in 3 GB RAM
• Multimodal — text, image, audio, and video input
• 140 languages — broadest language support in a mobile-class model
• Open weights — Gemma Terms of Use

For teams building edge AI agents, voice assistants, or mobile applications, Gemma 3n delivers multimodal capabilities that previously required cloud inference.

curl https://llm.hanzo.ai/v1/chat/completions \
  -H "Authorization: Bearer $HANZO_API_KEY" \
  -d '{"model": "google/gemma-3n-e4b-it", "messages": [{"role": "user", "content": "..."}]}'

Gemma 3 270M — The Smallest Gemma

Also released February 11, Gemma 3 270M is a 270M-parameter text-only model optimized for task-specific fine-tuning on constrained devices. It's the most power-efficient model in the Gemma family — an INT4-quantized version uses just 0.75% battery for 25 conversations on a Pixel 9 Pro. Weights are available on Hugging Face for local deployment.

100+ Models, One API

Gemma 3n joins the growing catalog on the Hanzo AI Gateway. Same endpoint, same API key, same billing — whether you're running a small on-device model or a 744B model in the cloud.

Browse the full catalog at hanzo.ai.