Get to know Gemma, friend Gemini.

You know, the Google home model besides Gemini has another Gemma example!

Get to know a model beyond Gemini like Gemma from the session "Evolution of Gemma 3n for Deploying Local Models on Every Device" by Witthawin Sripheanpol from Cloud Next Extended Bangkok 2025.

.

⭐ What's Gemma?

Gemini is an AI as a service to run web, made by DeepMind, a subsidiary company of Google that has made AI more and more accessible.

Gemma is an open-source model based on Gemini. Everyone can use it on their own machine without opening the internet! It is used on the local. There is no problem connecting the internet because it is on-device. The highlight is a small but smart model.

In the open source model line, there are DeepSeek, Qwen, Llama, Phi, about this.

Gemma also has many offspring, take it to fine-tune, develop various abilities such as MedGemma to LLM in medical matters, ShieldGemma to guardrail, can detect incoming messages as timeless or sensitive?

So what's the difference between Gemma 3 and Gemma 3n?

Gemma 3 provides multimodal, receives input, text, image, audio, video, and has sizing that is accessible. No need to find a server maran AI. Gemma 3n can be run on mobile, such as making apps or small devices like robotic, IoT.

.

⭐ Evolution from Gemma to Gemma3n.

If we see a strange AI name, how do we read it?

Gemma-3n-eXb-it

3 is the version, n is the Nano series, e is the effective parameter, X is the model size unit is a billion, it says is the model instruction-tuning? Is it trained as a conversation base? If it is, can it be used to do agent base?

The e was just memory efficiency with PLE saying that this model can split to which model to use. This makes the GPU use memory as we use the mobile RAM. Limit the small ROM, so some of it is processed.

Gemma understands 140 languages for text and 35 languages for multimodal and support Multimodal and Long-term conversation.

He said that Gemma 3n uses MobileNet-v5-300 and then we can take a picture, put a message, put a prompt and then understand those messages and vision quickly.

Advantages of Gemma 3n

- Optimized on-device performance: Good at working on-device and

- Privacy-first, offline-ready: privacy-first is offline, available now

- Multimodal understanding: Good understanding of letters, pictures, sounds, videos

- Dynamic resource usage: small, easy scale and also much cheaper.

.

⭐ How to use Gemma3n

There are now 4 versions. There are 2 base instructions on the top. Use in any chat section and 2 base models on the bottom.

- gemma-3n-E4B-it

- gemma-3n-E2B-it

- googlegemma-3n-E2B

- Gemma-3n-E4B

Can download all 4 locations: Hugging Face, kaggle, Ollama, LM Studio

And try to play in many places, such as Hugging Face, Google AI Studio. This session is done through Colab. There is a Python installation, load model and use it. If there is no text, it understands image and audio. Because there is a good encoder, it can match.

.

⭐ How to deploy Gemma3n on local device

Can I get an 8 GB model to 3 - 4 GB? It can be used for mobile app. If the RAM exceeds 8 GB with ONNX library supporting using the small model in the front-end, both website and mobile can be viewed on Hugging Face. There is a web front-end example. Use gemma-3n-E2B-it and support just the CPU. WebGPU wait.

They say that the Android side uses the onnx. I think the Hugging Face is more of a React Native tone because we don't use any npm.

The bigger the model, the more waste the screen card, and the more waste of money. If you move it, the CPU will reduce the cost by about 6 times, making the scale easier.

.

.

⭐, does it matter to Android Developer?

Of course, the Android Developer side also has other options, such as using the model that trains itself, Tensorflow Lite, MediaPipe, ML Kit, and Gemini Nano.

The Google AI for Developers document has a MediaPipe Solution, which has a lot of other solutions. The corresponding part is the LLM Inference API, the Generative AI that supports web, Android and iOS, and the customized model. The model used is Gemma-3 1B and Gemma-2 2B, an open-source model that can be downloaded through Hugging Face. [1]

The Android side can try to play at Google AI Edge Gallery what it does. [2] If you want to try it, try to create a new project and write code to download the model before using it in the app. [3]

As for ONNX, Teal, let's take a look at it and get confused. People don't touch Android NDK too. I'd rather use the usual method. 😆

.

What methods do friends use to implement AI in their projects?

# programmer # Developer # Includes IT matters # IT should know # googleai