Google's Gemma 4, released on 2 April 2026 under a fully permissive Apache 2.0 licence (Google Blog, MindStudio), is not just another open-weights refresh. It is the first open model family that credibly closes the gap with frontier proprietary systems while remaining deployable on hardware most technology teams already own. For CTOs evaluating whether to build or buy their AI capabilities, Gemma 4 shifts the calculus decisively.

Built from the same research underpinning Google's Gemini 3, the family spans four models — from edge devices drawing under 1.5 GB of memory to a flagship dense model ranking third among all open models globally on the Arena AI leaderboard (Google DeepMind). The Apache 2.0 licence removes every commercial restriction that constrained earlier Gemma releases, meaning firms can fine-tune on proprietary data and deploy commercially without licensing complexity (Beam AI).

Four models, from Raspberry Pi to data centre

The Gemma 4 family is designed around a simple insight: different business problems demand different compute budgets. Rather than shipping a single monolithic model, Google delivered four variants that share a common architecture but target radically different deployment envelopes (Labellerr, WaveSpeedAI).

Gemma 4 E2B packs 2.3 billion effective parameters into a footprint that runs on a Raspberry Pi 5 at 7.6 tokens per second. It handles text, images, audio, and video with a 128K-token context window — enough to process a short book in a single pass (MindStudio). Gemma 4 E4B doubles the effective parameter count to 4.5 billion, adding richer reasoning while still fitting comfortably on a laptop GPU. Both small models use Per-Layer Embeddings (PLE), a technique inherited from Gemma 3n that injects a lightweight conditioning signal at every transformer layer, squeezing more capability from fewer active parameters (Hugging Face).

Gemma 4 26B A4B introduces a Mixture-of-Experts (MoE) architecture with 128 fine-grained experts, of which only 8 (plus one oversized shared expert) activate per token (MindStudio). The result is a model with 25.2 billion total parameters but only 3.8 billion active at inference — delivering intelligence comparable to the 31B dense model at roughly the compute cost of a 4B model. At the top of the range, Gemma 4 31B is a conventional dense transformer with 30.7 billion parameters, a 256K context window, and benchmark scores that embarrass models five to twenty times its size (Gemma 4 Specs, NVIDIA NIM).

All four models share hybrid attention that alternates local sliding-window layers with global full-context layers, proportional RoPE for stable long-context reasoning, unified keys-and-values in global attention to halve KV-cache memory, and native functio