The new Gemma 3 artificial intelligence models from Google promise a significant impact with limited resources.

Google's artificial intelligence initiatives are closely tied to Gemini, which has been fundamentally integrated into its most popular products, encompassing both software and hardware in Worksuite. Additionally, the company has been launching various open-source AI models under the Gemma label for over a year. Recently, Google unveiled the third generation of its open-source artificial intelligence models with impressive claims.

The Gemma 3 models come in four variants, including 1 billion, 4 billion, 12 billion, and 27 billion parameters, and they are designed to operate on devices ranging from smartphones to powerful workstations. Google asserts that Gemma 3 is the best single-accelerator model in the world, meaning it can run on a single GPU or TPU unit without the need for a full cluster. This implies that a Gemma 3 AI model can natively run on the Tensor Processing Unit (TPU) of the Pixel smartphone, similarly to how the Gemini Nano model runs locally on phones.

One of the main advantages of Gemma 3 over the Gemini model family is that, being open-source, developers can package and adapt it to their specific needs within mobile and desktop applications. Another relevant aspect is that Gemma supports over 140 languages, of which 35 come in a pre-trained package. Furthermore, like the latest models from the Gemini 2.0 series, Gemma 3 has the ability to understand text, images, and videos, making it a multimodal model.

In terms of performance, Gemma 3 is claimed to outperform other popular open-source models, such as DeepSeek V3 and OpenAI's o3-mini model. Gemma 3 also offers a context window of 128,000 tokens, allowing it to process inputs the size of a 200-page book. For comparison, the context window of the Gemini 2.0 Flash Lite model is one million tokens. In AI models, an average English word is equivalent to approximately 1.3 tokens.

Moreover, Gemma 3 supports function calls and structured output, enabling it to interact with external datasets and perform tasks as an automated agent. This resembles how Gemini operates, facilitating seamless work across different platforms such as Gmail or Docs. Google's latest open-source AI models can be deployed locally or through their cloud-based platforms, such as the Vertex AI suite. The AI models of Gemma 3 are available on Google AI Studio, as well as on third-party repositories like Hugging Face, Ollama, and Kaggle.

Gemma 3 is part of a trend in the industry where companies develop Large Language Models (like Gemini in Google's case) while also releasing small language models (SLMs). Microsoft follows a similar strategy with its series of open-source small language models called Phi. Small language models like Gemma and Phi are extremely resource-efficient, making them ideal for operation on devices such as smartphones, and their low latency makes them particularly suitable for mobile applications.