Google releases Gemma 4 under Apache 2.0 — and that license change may matter more than benchmarks

Gemma 4 is part of Google DeepMind’s open model family and is released under the permissive Apache 2.0 license.

The release removes custom license friction: no custom clauses, no harmful‑use carve‑outs, and no redistribution restrictions.

Gemma 4 is organized into two deployment tiers: Workstation and Edge.

Workstation tier includes a 31B dense model and a 26B A4B Mixture‑of‑Experts model; both accept text and image inputs and offer 256K‑token context windows.

Edge tier includes compact E2B and E4B models for phones, embedded devices and laptops (text, image, audio; 128K context); E2B has 2.3B effective / 5.1B total parameters due to Per‑Layer Embeddings, while A4B is listed as 25.2B total with 3.8B active during inference (effective 26B); Google says competing open‑weight models like Qwen, Mistral and Arcee use similar permissive terms, with Alibaba’s Qwen3.5 Omni and Qwen 3.6 Plus cited as examples.