Gemma-3n-E4B-IT

Model Description

Gemma 3n E4B-IT, developed by Google DeepMind, is a 4-billion-parameter efficient multimodal model.
Built with MatFormer architecture and dynamic parameter activation, it delivers strong text, image, audio, and video understanding while remaining lightweight enough for on-device deployment.
It supports a 32K context window and multilingual inputs across more than 140 languages.

Features

• Multimodal input: text, image (up to 768×768), audio, and video.
• Efficient design: parameter skipping and caching enable deployment on edge devices.
• Large context window: up to 32K tokens.
• Multilingual: trained on 140+ languages.
• Compact but strong: achieves benchmark scores competitive with much larger models.

Use Cases

• Visual question answering and captioning
• Conversational agents with multimodal inputs
• On-device assistants for mobile and embedded systems
• Multilingual summarization, translation, and transcription

Inputs and Outputs

Input:

• Text prompts or dialogue
• Single image (tokenized for processing)
• Multiple image inputs and audio inputs support coming soon!

Output:

• Generated text (answers, captions, translations, summaries)

---

How to use

1) Install Nexa-SDK

Download and follow the steps under "Deploy Section" Nexa's model page: Download Windows SDK

2) Get an access token

Create a token in the Model Hub, then log in:

nexa config set license '<access_token>'

3) Run the model

Running: bash nexa infer NexaAI/gemma-3n

License

• Licensed under Google’s Gemma terms of use. See Hugging Face model card for details.

References

• Hugging Face: google/gemma-3n-E4B-it
• Gemma 3n documentation
• Google AI blog announcement