Gemma 3n E4B

Name: Gemma 3n E4B
Author: Google

Google · Gemma 3n · released 2025-06-26 · data checked 2026-11-02

Gemma 3n E4B is a 7.8B-parameter dense model from Google (United States), released 2025-06-26 under the Gemma license. It scores 18.2 on the Small LM Index, #25 of 26 small models. It needs about 5.5 GB of memory at 4-bit with an 8K-token context.

Mobile-first architecture (MatFormer) with selective parameter loading.

Small LM Index

18.2

#25 of 26 scored small models

Likely 15.8–21.0 · from 4 of 6 core benchmarks

Parameters

7.8B

4B effective

Price per 1M tokens

—

No hosted API on OpenRouter; self-host

Context window

32K

tokens

Benchmarks

Each bar shows this model's score. The tick marks the best small model on that benchmark. Sources: 6 reported by Google. How we label sources.

MMLU-Pro Lab-reported Knowledge · Instruct model, 0-shot (all scores)50.6

Class best: Qwen3.5 9B 82.5

GPQA Diamond Lab-reported Knowledge · GPQA Diamond23.7

Class best: Qwen3.5 9B 81.7

AIME Lab-reported Math · AIME 202511.6

Class best: Apriel 1.6 15B Thinker 88.0

LiveCodeBench Lab-reported Coding · LiveCodeBench v525.7

Class best: Granite 4.2 8B 73.2

MMLU Lab-reported Knowledge64.9

Class best: OLMo 3 7B Think 77.8

Multilingual MMLU Lab-reported Multilingual · Global-MMLU60.3

Not reported: estimated

What the index model expects from this model's other scores. Not a measurement.

Humanity's Last Exam≈ 1

SWE-bench Verified≈ 5

Specifications

Lab: Google (US · United States)
Architecture: Dense transformer
Total parameters: 7.8B
Active parameters: All (dense)
Size class: Small
Context window: 32K tokens
Input modalities: text, image, audio, video
Reasoning: Instruct (no thinking)
License: Gemma
Openness: Open weights
Memory to run: ~5.5 GB at 4-bit, 8K context (more sizes)
Hugging Face: 12.7K downloads (30 days) · 936 likes

Weights on Hugging Face ↗

Compare with

Gemma 3n E4B vs Falcon H1R 7B (7.6B, index 57.5)
Gemma 3n E4B vs Ling 3.0 Tiny (7.9B, index 51.9)
Gemma 3n E4B vs Edge0 8B-A1B (preview) (7.9B, index 47.2)
Gemma 3n E4B vs MiMo-7B-RL (7.8B, index 40.7)
Gemma 3n E4B vs DeepSeek-R1-Distill-Qwen-7B (7.6B, index 39.8)
Gemma 3n E4B vs Granite 4.2 8B (8B, index 63.5)

Sources

https://huggingface.co/google/gemma-3n-E4B-it

Something wrong here? Report it.

Run it locally

Uses unsloth/gemma-3n-E4B-it-GGUF (Q4_K_M). Swap the tag for another size: F16, IQ4_NL, IQ4_XS, Q2_K, Q2_K_L, Q3_K_M, Q3_K_S, Q4_0….

Ollama

ollama run hf.co/unsloth/gemma-3n-E4B-it-GGUF:Q4_K_M

llama.cpp

llama-server -hf unsloth/gemma-3n-E4B-it-GGUF:Q4_K_M

Then open http://localhost:8080 for a chat window.

MLX (Mac)

pip install -U mlx-lm
mlx_lm.chat --model mlx-community/gemma-3n-E4B-it-lm-4bit

Text only; for images use mlx-vlm.

vLLM (GPU server)

vllm serve google/gemma-3n-E4B-it

Full-precision weights, for a server GPU.

Open in LM Studio

Memory needed: weights, context cache and runtime buffers.

Quantization	4K context	32K context
16-bit F16	16 GB	17 GB
8-bit Q8_0	9.1 GB	9.3 GB
4-bit Q4_K_M	5.5 GB	5.7 GB

Check it against your GPU, Mac or phone →