Google's new open model 'Gemma 4' features outstanding reasoning capabilities. Notably, the E2B model is light enough to run on just 700 lines of C code, making it suitable for various devices like smartphones.
Imagine this: you wake up in the morning and say to your smartphone, “Organize my meeting schedule for today and list them by priority.” Previously, this request would have flown across the internet to Google’s massive data centers to undergo complex calculations before returning. Now, the entire process is handled in an instant right on your smartphone. The star of this show is Google’s latest, ambitiously released artificial intelligence model, ‘Gemma 4.’
Why does this matter?
Most of the powerful AIs we have been using until now required an internet connection. This is because the ‘parameters’ (adjustable numerical values inside the model)—the brain of the AI model—were too enormous to fit on personal devices. However, Gemma 4 is changing the game.
Gemma 4 demonstrates a surprising level of ‘intelligence relative to parameters’ and is optimized for complex reasoning and AI agent (AI that performs tasks on the user’s behalf) work Source: Gemma 4: Our most capable open models to date Source: Gemma 4 - Google DeepMind. This means that high-level task assistance becomes possible on your mobile phone without an internet connection.
Easy to understand: The magic of an ultra-compact guidebook
What is the secret that allows Gemma 4 to run on a smartphone? The key is ‘efficiency.’ The ‘E2B’ model, the smallest in the Gemma 4 series, was designed to run on just 700 lines of C code Source: Gemma 4 E2B inference 700 lines code.
To use an easy analogy: if a conventional massive AI model were a team that required 100 experts to gather and discuss before reaching a conclusion, the Gemma 4 E2B is like a veteran carrying an ‘ultra-compact guidebook’ that contains only the essential know-how of those experts. Because the guidebook is thin, it can naturally judge situations and provide answers quickly with fewer resources.
Furthermore, Google has added a magical optimization technique called ‘Multi-token prediction’ Source: Google’s multi-token prediction. This is similar to a writer writing a book where an assistant sitting next to them suggests upcoming sentences in advance, and the writer quickly checks if the suggestions are correct. By having a small model (an auxiliary model) propose several tokens (data fragments that AI splits when processing language) in advance and having the main model verify them at once, inference speed has been dramatically increased Source: Google’s multi-token prediction.
How far have we come?
Gemma 4 is not just a model that writes text well. These models support ‘Multimodal’ capabilities (the ability to simultaneously understand not just text, but also various forms of data like images and audio) Source: Gemma 4 model overview Source: Gemma 4. Currently, Gemma 4 has been released in various sizes—including E2B, E4B, 12B, 31B, and 26B A4B—to match the performance and purpose of the user’s device Source: Gemma 4 model overview.
Developers and users are already actively utilizing it through various platforms such as Google AI Studio, Vertex AI, Hugging Face, and Ollama, and you can run it immediately on your personal computer or laptop through popular inference frameworks such as llama.cpp and vLLM Source: Gemma 4 - Google DeepMind.
Future changes
Gemma 4 is the first step toward the daily integration of AI. Moving forward, appliances, cars, and mobile phones equipped with high-efficiency models like Gemma 4 will evolve from passive tools waiting for commands into true ‘agents’ that understand situations and solve problems on behalf of the user. Above all, since powerful AI features can be enjoyed without sending personal data outside the device, privacy concerns are also expected to be significantly improved.
References
-
[Gemma 4 E2B inference in 700 lines of C Modern Orange](https://modernorange.io/item/49468286) - Welcome Gemma 4: Frontier multimodal intelligence on device
- Gemma 4 — Google DeepMind
- Google says multi-token prediction makes Gemma 4 up to… - YouTube
- Gemma 4: Our most capable open models to date
-
[Gemma 4 model overview Google AI for Developers](https://ai.google.dev/gemma/docs/core) - Gemma 4: Our most capable open models to date
- Gemma 4: Our most capable open models to date
- Gemma 4: Our most capable open models to date
- Gemma 4: Our most capable open models to date
-
[Gemma 4 12B: обзор локальной мультимодальной… AiManual](https://ai-manual.ru/article/gemma-4-12b-pervoe-ruchnoe-testirovanie-lokalnoj-multimodalnoj-modeli-s-zreniem-audio-i-vyizovom-instrumentov/) - Gemma 4
- It can only process text
- It is optimized for advanced reasoning and agent tasks
- It is so heavy that it only works on supercomputers
- It requires a million lines of Python code
- It is capable of inference with just 700 lines of C code
- It is 100 times slower than existing models
- It increases training time
- It enhances security
- It speeds up inference by having the main model verify multiple tokens proposed by a draft model at once