gemma-4-E4B-it

Gemma 4 E4B IT is the instruction-tuned release of the larger of the two edge-sized Gemma 4 models, published by Google DeepMind on April 2, 2026 under the Apache 2.0 license. It runs with 4.5 billion effective parameters, 8 billion once embeddings are counted, and uses Per-Layer Embeddings to keep the working footprint small enough for high-end phones, laptops, and edge devices. Its 42 layers use 512-token sliding windows, and it carries a 128K token context window. Unlike the 31B model, E4B accepts audio as well as text and images, pairing a roughly 300M parameter audio encoder with a roughly 150M parameter vision encoder, and it handles speech recognition and speech-to-translated-text natively. Thinking mode is switched with dedicated control tokens and function calling is built in, so it can act as a small on-device agent. Google reports 69.4% on MMLU Pro, 58.6% on GPQA Diamond, 52.0% on LiveCodeBench v6, and 42.5% on AIME 2026 for the instruction-tuned model. E4B IT fits assistants that must run offline or under tight memory budgets, voice-driven apps, and multilingual chat where the model's pretraining on more than 140 languages matters more than peak reasoning depth.