Gemma 4 E4B

Gemma 4 E4B is the larger of the two on-device models of the family, and it reached Hugging Face on 2 March 2026 - nine days before the 26B and 31B models that the launch was built around. The E stands for effective. The model stores about 8 billion parameters, but only some 4.5 billion of them are ever active in memory, because it uses Per-Layer Embeddings: instead of stacking more layers, each decoder layer gets its own small embedding for every token. Those embedding tables are large but serve only as lookups, so the memory a phone actually has to hold is far smaller than the download suggests. It has 42 layers, a 128,000-token context, a 262,000-token vocabulary and 512-token sliding windows - half the window of the bigger models, another concession to memory. Modalities are unusually generous for the size: text, image and audio in, text out, through a vision encoder of roughly 150 million parameters and an audio encoder of roughly 300 million. Speech recognition and speech-to-translated-text are part of the package. What this size buys and what it costs are both visible in the numbers. Against Gemma 3 27B, a model more than three times larger from ten months earlier, E4B is ahead almost everywhere: 69.4 percent against 67.6 on MMLU Pro, 42.5 against 20.8 on AIME 2026, 52.0 against 29.1 on LiveCodeBench, 58.6 against 42.4 on GPQA Diamond. Against its own 31B sibling it is far behind - 42.5 against 89.2 on AIME - which is the honest shape of the trade. Read on 25 August 2026 the instruction-tuned repository records 5.21 million downloads in thirty days, the third highest in the family.

▸Videos

No videos yet.

▸Reviews