GOOGLE / GEMMA
THE SHORT VERSION

The launch offers four sizes, with support for reasoning, tool calls and multimodal inputs.

What happened

The lineup includes E2B, E4B, a 26B mixture-of-experts model and a 31B dense model. Google describes image and video input across the family, with native audio input in the two smaller variants.

Why it matters

Different sizes let developers match capability to memory, latency and deployment needs. Local operation can also make an application less dependent on sending every request to a hosted service.

The bigger picture

Our take. Choose a variant using actual hardware and task measurements. Google's intelligence-per-parameter and leaderboard claims are useful context, but do not specify the full memory footprint, operating cost or reliability of a finished application.

Primary source: Read the original: Google · Gemma 4 launch announcement ↗

Source checked November 18, 2026. Company claims are attributed; paragraphs marked “Our take” are editorial interpretation.

UNDERSTAND THE IDEA BEHIND THE NEWSUnderstand the units behind model input and output↗