Gemma 3n E4B

Gemma 3n E4B is Google's mobile-first multimodal model, generally available June 26, 2025 (previewed at Google I/O, May 20, 2025). "E4B" = effective-4B: the checkpoint totals ~8B parameters but uses the MatFormer (Matryoshka Transformer) nested architecture, so a fully functional 4B submodel runs inside the larger model and memory footprint stays around 3GB (E2B sibling: ~2GB). It natively handles text, image, audio, and video input with a 32K-token context, 140-language text support, and on-device speech-to-text/translation via a Universal Speech Model-based audio encoder (30-second clips) plus a MobileNet-V5 vision encoder (provider-claimed 60 FPS video on Pixel).

Benchmarks are provider-reported (official model card): LMArena >1300 — the first model under 10B parameters to break 1300; MMLU 64.9, MBPP 63.6, HumanEval 75.0, Global-MMLU-Lite 64.5, MMLU-Pro 50.6 (E4B IT). Per-layer embeddings cut parameters ~46% and KV cache sharing doubles prefill speed vs Gemma 3 4B (provider claims).

Licensed under the Gemma Terms of Use (custom, non-OSI; gated download requiring terms acceptance; commercial use allowed with restrictions). No paid official API — available free in Google AI Studio, on Kaggle, and self-hosted via Transformers, llama.cpp, Ollama, MLX, and Google AI Edge. The reference choice for private, offline, on-device multimodal AI.

Provider-reported (official model card / Google): LMArena >1300 (first sub-10B model to break 1300); MMLU 64.9, MBPP 63.6, HumanEval 75.0, Global-MMLU-Lite 64.5, MMLU-Pro 50.6 (E4B IT); 2x prefill latency improvement over Gemma 3 4B.

Independent: Not yet independently replicated at batch time; launch press (Neowin, Campus Technology, Jul 2025) confirmed the LMArena claim as Google's reported figure.