Gemma 4 E2B

Gemma 4 E2B is an open multimodal language model developed by Google DeepMind as part of the Gemma 4 family (released 2 April 2026). The "E" stands for "effective" parameters: the model has 2.3B effective parameters (5.1B including embeddings) thanks to Per-Layer Embeddings (PLE), which reduces memory use during on-device inference.

It accepts text, image and audio as input and generates text. It supports a 128K-token context window, over 140 languages, native function calling and configurable thinking (reasoning) modes. The architecture is dense and uses a hybrid attention mechanism interleaving local sliding-window attention with full global attention, plus Proportional RoPE.

Weights are released under the Apache 2.0 license in pre-trained and instruction-tuned variants; the model can be downloaded (Hugging Face) and run locally or on-device, and is also available through Google AI Studio. The training-data knowledge cutoff is January 2025.