Model Details
Gemma 3n E2B
Quick answer: Gemma 3n E2B is Google's smallest May 2025 on-device multimodal model with 2B effective parameters, scoring 51.7% on ARC-C, 40.5% on MMLU-Pro, and 72.2% on HellaSwag. It is designed for highly constrained on-device deployment with text, image, and audio input support.
At a Glance
Where Gemma 3n E2B leads
- Smallest Gemma 3n model — lowest inference cost
- Multimodal: text, image, and audio inputs at 2B effective param scale
- MatFormer nested architecture for efficient on-device execution
- Suitable for most-constrained mobile hardware
Where it lags
- 40.5% MMLU-Pro — limited academic knowledge at 2B scale
- Below Gemma 3n E4B on all benchmarks (expected at 2B vs 4B)
- Text-only alternatives (Llama 3.2 1B) offer similar size
Best for: Most-constrained on-device deployment; wearables and IoT; minimum-footprint multimodal AI.
What Gemma 3n E2B Is
Gemma 3n E2B is the smallest model in Google's Gemma 3n family, announced at Google I/O 2025. At 2B effective parameters with the MatFormer nested architecture, it targets the most resource-constrained on-device use cases — wearables, IoT, and entry-level smartphones.
Like Gemma 3n E4B, it supports text, image, and audio inputs — multimodal capability at scales where most models are text-only. The MatFormer architecture allows dynamic compute allocation, so actual inference cost varies by task complexity.
For most mobile applications, Gemma 3n E4B provides meaningfully better benchmarks (50.6% vs 40.5% MMLU-Pro) with acceptable additional compute. E2B is preferred when E4B exceeds device constraints.
Specifications
| Field | Value |
|---|---|
| Organization | |
| Effective parameters | 2B (MatFormer nested) |
| License | Gemma Terms of Use |
| HuggingFace | google/gemma-3n-E2B-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 | 51.7% | Benchgen evaluation | 2025-07 |
| HellaSwag | 72.2% | Benchgen evaluation | 2025-07 |
| MMLU-Pro | 40.5% | Benchgen evaluation | 2025-07 |
Gemma 3n E2B vs Alternatives
| Model | MMLU-Pro | Modality | Effective Params |
|---|---|---|---|
| Gemma 3n E2B | 40.5% | Text+Vision+Audio | 2B |
| Gemma 3n E4B | 50.6% | Text+Vision+Audio | 4B |
| Llama 3.2 3B Instruct | — | Text only | 3B |
Gemma 3n E4B provides +10pp MMLU-Pro at 2× the effective parameters. Use E2B only when E4B won't fit on the target device.
Related Models & Benchmarks
- Compare: Gemma 3n E4B · Gemma 3 12B · Llama 3.2 3B Instruct
- Benchmarks: MMLU-Pro · HellaSwag
Frequently Asked Questions
What is Gemma 3n E2B?
Google's May 2025 smallest on-device multimodal model with 2B effective parameters (MatFormer), scoring 51.7% ARC-C, 40.5% MMLU-Pro, and 72.2% HellaSwag. Supports text, image, and audio.Should I use Gemma 3n E2B or E4B?
E4B for better performance (50.6% vs 40.5% MMLU-Pro). Use E2B only when E4B exceeds device hardware constraints.Specs from Google's official Gemma 3n release (May 2025) and Benchgen evaluations. Last updated 2026-07-24.