Gemma 4 E2B

Gemma 4 E2B is the smaller edge tier of Google's Gemma 4 family, released late March 2026 under the Apache 2.0 license — the first Gemma generation to use a fully permissive license rather than Gemma Terms of Use. It carries ~2.3B effective parameters (5.1B stored, using per-layer embeddings), a 128K context window, and natively handles text, image, and audio input with a redesigned speech encoder trained on 140+ languages. Google reported large gains on STEM, multimodal, and long-context benchmarks plus major safety-category improvements over Gemma 3n, with explicit thinking-mode support via a chat-role token.\n\nThe honest picture: E2B is genuinely the most capable open model at its size in 2026, but the effective-vs-stored parameter split (roughly 2.3B running on 5.1B of stored weights) can surprise people budgeting disk or RAM. It also drops MatFormer nesting from Gemma 3n, so you pick E2B or E4B rather than slicing your own size.