Gemma 4 (E2B, E4B, 12B, 26B A4B, 31B)
Gemma 4 is Google's current family of open-weight models, released in April 2026 under Apache 2.0. It ranges from small E2B/E4B models for phones to a 31B dense model, all reading text and images and writing text. A 12B 'unified'…
Best for
Self-hosted or on-device AI where you need an open, commercially permissive model in sizes from phone to single-GPU server.
Overview
Gemma 4 is Google's current family of open-weight models, released in April 2026 under Apache 2.0. It ranges from small E2B/E4B models for phones to a 31B dense model, all reading text and images and writing text. A 12B 'unified' model and quantized (QAT) versions followed in June 2026.
Strengths and limits
- Apache 2.0 license (more permissive than earlier Gemma terms)
- Wide size range, incl. E2B/E4B for mobile and a 26B mixture-of-experts
- Strong reported scores for its size (31B: 85.2% MMLU Pro)
- Audio input only on E2B, E4B and 12B; video handled as frame sequences
- Well behind frontier closed models on hard agentic tasks
Published benchmark scores
Scores are as published by the source; settings (reasoning effort, tools, attempts) differ between labs, so compare with care. How to read benchmarks →