Gemma 4 E2B IT

Gemma 4 E2B IT sits at the smallest end of the Gemma 4 family released by Google DeepMind, a multimodal lineup that pairs dense and mixture-of-experts architectures across five size tiers. The instruction-tuned variant is distributed with open weights under an Apache 2.0 license, letting teams embed it directly in commercial products. Its design leans on Per-Layer Embeddings, the same efficiency technique used by the E4B sibling, to keep memory demands low enough for laptops and high-end phones. The card positions the family as strong reasoners with configurable thinking modes, and E2B inherits that reasoning-oriented posture rather than treating it as a bolt-on feature.

On input, the model accepts text, images, and audio, then generates text output, making it well suited to workflows that mix document understanding, OCR, voice cues, and lightweight agentic loops. Independent hosting documentation reports roughly 5.1 billion total parameters with about 2.3 billion active during inference, 35 layers, and a hybrid attention pattern that uses a 512-token sliding window, which together support responsive local latency. Developers highlight the E2B variant as a fit for routine agent tasks, on-device assistants, and OCR pipelines where running on a smartphone-class footprint matters more than chasing the largest context. Combined with open weights and broad language coverage in the wider family, it offers a pragmatic bridge between quality and deployability for production teams that need a private, self-hosted model.