Model Details
Gemma 3n E4B
Quick answer: Gemma 3n E4B is Google's May 2025 on-device multimodal model with 4B effective parameters, scoring 61.6% on ARC-C, 75.0% on HumanEval, and 50.6% on MMLU-Pro. The "3n" architecture uses MatFormer nested structures for efficient on-device deployment while supporting text, image, and audio inputs.
At a Glance
Where Gemma 3n E4B leads
- 75.0% HumanEval — excellent code completion at 4B effective params
- Multimodal: text, image, and audio inputs on-device
- MatFormer nested architecture — efficient on-device inference
- Gemma Terms of Use with on-device distribution rights
- Google I/O 2025 showcase model
Where it lags
- 4B effective params: below larger models on all benchmarks
- 50.6% MMLU-Pro — moderate knowledge
- On-device focus limits server-side competitive comparison
- Gemma Terms of Use (not Apache 2.0)
Best for: On-device multimodal applications; mobile AI with image + audio support; Google Pixel and Android on-device deployments.
What Gemma 3n E4B Is
Gemma 3n E4B (released May 20, 2025) is Google's on-device multimodal model line, announced at Google I/O 2025. The "3n" designation refers to the new MatFormer-based nested architecture, designed for efficient on-device deployment on mobile and edge hardware.
Unlike previous Gemma models (1B, 4B, etc. which have constant parameter counts), Gemma 3n E4B refers to "4 billion effective parameters" — the MatFormer architecture allows the model to dynamically adjust compute based on task complexity.
The model supports text, image, and audio inputs — making it the first Gemma model with audio understanding. This multimodal breadth at the 4B effective parameter scale is its primary differentiator from Llama 3.2 3B (text-only) and similar small models.
Specifications
| Field | Value |
|---|---|
| Organization | |
| Effective parameters | 4B (MatFormer nested) |
| License | Gemma Terms of Use |
| HuggingFace | google/gemma-3n-E4B-it |
| Release date | May 20, 2025 |
| Knowledge cutoff | February 2025 |
| Modality | Text + Vision + Audio |
| Architecture | MatFormer (nested) |
Pricing
Open weights under Gemma Terms of Use — free to self-host. Optimised for on-device deployment.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| ARC-C | 61.6% | Benchgen evaluation | 2025-07 |
| HellaSwag | 78.6% | Benchgen evaluation | 2025-07 |
| HumanEval | 75.0% | Benchgen evaluation | 2025-07 |
| MMLU-Pro | 50.6% | Benchgen evaluation | 2025-07 |
Gemma 3n E4B vs Alternatives
| Model | HumanEval | MMLU-Pro | Modality | Target |
|---|---|---|---|---|
| Gemma 3n E4B | 75.0% | 50.6% | Text+Vision+Audio | On-device |
| Gemma 3n E2B | — | 40.5% | Text+Vision+Audio | On-device |
| Llama 3.2 3B Instruct | — | — | Text | On-device |
| Gemma 3 12B | 85.4% | 60.6% | Text+Vision | Server |
Gemma 3n E4B vs Llama 3.2 3B: adds image and audio inputs at comparable parameter count. vs Gemma 3n E2B: higher accuracy (50.6% vs 40.5% MMLU-Pro) at higher compute cost.
Related Models & Benchmarks
- Compare: Gemma 3n E2B · Gemma 3 12B · Llama 3.2 3B Instruct
- Benchmarks: HumanEval · MMLU-Pro
Frequently Asked Questions
What is Gemma 3n E4B?
Google's May 2025 on-device multimodal model with 4B effective parameters (MatFormer), scoring 75.0% HumanEval and 50.6% MMLU-Pro. Supports text, image, and audio inputs.What does "E4B" mean in Gemma 3n?
E4B = 4 billion effective parameters. The MatFormer nested architecture uses dynamic compute allocation — "effective" refers to the active parameter count during inference.Specs from Google's official Gemma 3n release at Google I/O (May 2025) and Benchgen evaluations. Last updated 2026-07-24.