gemma-3n-E2B-it

Gemma 3n E2B IT is Google DeepMind's instruction-tuned, open-weights multimodal model optimized for efficient on-device and edge deployment. Built on the MatFormer (Matryoshka Transformer) architecture, it nests a fully functional E2B sub-model inside the larger E4B model, enabling selective parameter activation and Per-Layer Embeddings (PLE) that reduce accelerator memory to roughly 2GB effective footprint despite ~6B total parameters. The model accepts text, image, video, and audio inputs and generates text outputs, with support for 140+ languages and a 32K-token context window.

Architecturally, Gemma 3n extends the Gemma family with innovations including Alternating Updates (AltUp), Learned Augmented Residual Layer (LAuReL), activation sparsity, and KV cache sharing. It uses a MobileNet v5 vision encoder (768×768 default) and a Universal Speech Model (USM)-based audio encoder alongside a text backbone with alternating local sliding-window and global self-attention layers. The IT variant was post-trained with knowledge distillation and reinforcement learning on approximately 11 trillion tokens with a June 2024 knowledge cutoff.

On benchmarks, E2B IT achieves competitive results for its size class: 60.1 on MMLU, 40.5 on MMLU-Pro, 66.5 HumanEval pass@1, and 24.8 on GPQA Diamond. Released June 26, 2025 under the Gemma Terms of Use, it is available on Hugging Face, Kaggle, and Google AI Edge, with MatFormer Lab enabling custom Mix-n-Match sizes between E2B and E4B for hardware-specific tuning.