The newest generation of Google’s open AI models, known as Gemma 3n, is a big improvement over the previous models. The entire version is now available and prepared to run straight on your hardware, following a peek at Google I/O last month.
Gemma is a family of open AI models, in case you were unaware. In contrast to Gemini, which is Google’s closed, proprietary powerhouse, it is made for developers to download and alter.
The model has advanced from being a text-based model to one that can now natively interpret inputs including photos, audio, and video to produce text. It is said to perform better at things like reasoning and coding, and it can operate on devices with as little as 2GB of RAM.
The complete list of enhancements as stated by Google is as follows:
- Multimodal by design:
Image, audio, video, and text inputs and outputs are all natively supported by Gemma 3n.
- Optimized for on-device:
Gemma 3n versions, which were designed with efficiency in mind, come in two sizes, E2B and E4B, depending on the effective parameters. Despite having 5B and 8B of raw parameters, respectively, they can operate with as little as 2GB (E2B) and 3GB (E4B) of memory thanks to architectural advancements that make their memory footprint comparable to that of conventional 2B and 4B models.
- Groundbreaking architecture:
Basically, Gemma 3n includes new elements such as Per Layer Embeddings (PLE) for memory efficiency, the MatFormer architecture for computational flexibility, and new audio and MobileNet-v5 based vision encoders tailored for on-device use cases.
- Enhanced quality:
Gemma 3n offers quality enhancements in math, coding, reasoning, and multilinguality (supporting 140 languages for text and multimodal understanding of 35 languages).
A new architecture is at the heart of its effectiveness. MatFormer is called by Google. Google compares it to a Russian Matryoshka doll, which has a miniature, fully functional version within a larger model.
This makes it possible for a single model to function at various sizes for various activities. Regarding benchmarks, the larger E4B model is the first to surpass an LMArena score of 1300 when operating under 10B parameters.
More details can be found in the official announcement post.
