Open SourceGoogle DeepMindReleased 2025-03

Gemma 3

Google's open-weights model family optimized for on-device and edge deployment

Context window

128K

Input / 1M tokens

Free

Output / 1M tokens

Free

Provider

Google DeepMind

Free to download and self-host under Gemma Terms of Use. Available on Google AI Studio for free prototyping.

Gemma 3 comes in 1B, 4B, 12B, and 27B parameter sizes, with the 27B version competitive with models twice its size. Unlike Llama or Qwen, Gemma 3 is explicitly tuned for deployment on consumer hardware - the 4B model runs comfortably on a MacBook Pro. It's also natively multimodal, accepting image inputs in addition to text.

Strengths

  • Runs on consumer hardware (4B and 12B variants)
  • Multimodal input support
  • Strong benchmark performance relative to size
  • Tight Keras and JAX integration
  • Good instruction following out of the box

Best for developers who...

On-device and edge inferenceLow-resource environmentsPrototyping with free Google AI Studio accessResearchers benchmarking small models

Benchmarks

BenchmarkScoreNotes
MMLU~76%27B variant; strong for open-weights at this size

Source: Google Gemma 3 announcement

Compare Gemma 3 with

All model comparisons →

Learn the concepts