Gemma 4 E4B

Gemma 4 E4B is the larger of Google's two edge-optimized Gemma 4 variants, released March 31, 2026 alongside E2B, the 26B-A4B MoE, and the 31B dense flagship. It carries ~4B effective parameters (8B stored with per-layer embeddings), a 128K context window, and native text, image, and audio input via a compact conformer speech encoder — smaller and more responsive than Gemma 3n's, with 40ms frames and 25 tokens/second of audio. Like the whole Gemma 4 family, it ships under the Apache 2.0 license, with a January 2025 knowledge cutoff and safety improvements across text and vision modalities.\n\nThe honest picture: E4B trades roughly double the memory of E2B for noticeably better STEM and multimodal scores (its GPQA-Diamond jumped to 58.6% from 23.7% on Gemma 3n E4B, per Google's reported comparisons). But like E2B it no longer offers Gemma 3n's MatFormer self-sizing — the size tiers are fixed. And 'open' here still excludes the audio-encoder training data details.