NemoClaw Knowledge Wiki

Google Gemma 4: Efficient 2.3B Parameter Multimodal Edge AI

4 min read

Google Gemma 4: Efficient 2.3B Parameter Multimodal Edge AI

Generated: 2026-04-22 · API: Gemini 2.5 Flash · Modes: Summary

Google Gemma 4: Efficient 2.3B Parameter Multimodal Edge AI

Clip title: The 2.3B AI Model that “Thinks” like a 70B (Gemma 4) Author / channel: Better Stack URL: https://www.youtube.com/watch?v=ZxQ2DuejRhU

Summary

Google recently unveiled Gemma 4, a new family of open-source language models released under the permissive Apache 2.0 license. This video explores the capabilities of Gemma 4, particularly its smaller “edge versions” (E2B and E4B models), which are designed to run efficiently and entirely offline on various devices, from smartphones and Android flagships to Raspberry Pis. The presenter highlights the increasing competition in developing highly intelligent, compact models and sets out to test Gemma 4 against a previous model (Qwen 3.5) across several practical scenarios.

A key innovation behind Gemma 4’s efficiency is what Google calls “Per-Layer Embeddings” (PLE). Unlike traditional transformer models where a token receives a single embedding at the start, Gemma 4 assigns a unique set of embeddings to each layer. This allows the model to introduce new information precisely when needed, resulting in high “intelligence density.” For instance, the E2B model, despite having a reasoning depth equivalent to a 5 billion parameter model, only utilizes about 2.3 billion active parameters during inference, requiring less than 1.5GB of RAM. Beyond text, Gemma 4 is natively multimodal, processing vision, text, and audio within a unified architecture. It also features a “Thinking Mode” that uses an internal reasoning chain to verify its logic, preventing common errors found in smaller models, and boasts a large context window and support for over 140 languages. Benchmarks indicate impressive performance, with the E4B model achieving more than double the score of larger previous-generation models on complex math challenges and showing significant improvement in tool use accuracy via “Agent Skills.”

The video then delves into practical tests using the E2B and E4B models running locally via LM Studio and Cline, offline. In a coding task to generate a cafe website (HTML, CSS, JavaScript), the E2B model delivered underwhelming results, taking 1.5 minutes but producing incomplete code with non-functional elements and an empty JavaScript file. The E4B model, while slower at 3.5 minutes, produced a noticeably better and more functional website, including a working shopping cart, which its smaller counterparts had failed to do. However, both models still produced a visually basic design, leading the presenter to conclude that these small models are not yet suitable for complex or production-level coding tasks.

Further testing on an iPhone using Google’s AI Edge Gallery app demonstrated Gemma 4’s performance on edge devices. The E2B model responded quickly to text prompts, providing a detailed, albeit somewhat verbose, response to a car wash dilemma. For image understanding, it correctly identified a dog and its characteristics but misidentified its breed. In an OCR test with Latvian text, Gemma 4 successfully identified the language and translated most of the content accurately, despite some grammatical oddities. It also handled a basic conversation in Latvian, showing impressive multilingual knowledge for its size, confirming its knowledge cutoff was January 2025.

In conclusion, Gemma 4 appears to be a highly capable open-source model that largely lives up to its advertised features, particularly its multimodal and reasoning abilities within a compact footprint. While it may lack creativity in web design, its performance in tasks like OCR and basic logical reasoning on edge devices is remarkable for its size. A current limitation is the lack of official MLX bindings for local iOS development, forcing reliance on Google’s own app, though community projects like SwiftLM are emerging. Overall, Gemma 4 represents a significant advancement in small, on-device AI, proving that such models can handle complex tasks successfully.

Graph View