Google’s Gemma 4 AI Runs on Old Consoles

Google’s Gemma 4 AI Runs on Old Consoles

Google DeepMind has released a new family of AI models called Gemma. These models are free and open, meaning anyone can use them.

This is a big deal because many powerful AI tools are expensive or require a subscription. Some users have even lost access to AI services they paid for due to high demand.

The smallest Gemma models are designed to be very lightweight. They only need a few gigabytes of memory to run. This means they don’t require expensive graphics cards, or GPUs.

You can even run them on your phone without an internet connection. People are already using Gemma to build apps for offline translation and summarizing text.

Surprising Power in Small Packages

The smallest Gemma model is surprisingly capable. It can even run on an original Nintendo Switch, a gaming console with limited processing power.

This shows how efficient the Gemma models are. They can run the 2 billion parameter Gemma 4 model on older hardware.

Google DeepMind also offers a larger 31 billion parameter Gemma model. This model performs exceptionally well, even when compared to AI models that are ten times its size. This is impressive because larger models usually require more computing power.

How Gemma Achieves Efficiency

One reason for Gemma’s efficiency is its training data. Google used highly curated data, filtering out a lot of noise.

This is similar to how humans benefit from focusing on good information. The model also uses a new type of attention mechanism called ‘hybrid attention’.

Hybrid attention combines two ways of processing information. It uses ‘sliding window’ attention for detailed, line-by-line processing, like reading a book.

It also uses ‘global attention’ to understand the bigger picture, like knowing which book you are reading. This helps the AI understand context better.

Better Image Understanding and Memory

Gemma 4 also shows improvements in understanding images. Unlike previous versions, it doesn’t force images into a square shape, which can lose detail.

It processes images as they are, leading to better results on image-related tasks. This is a significant upgrade from earlier models.

The model also uses a ‘shared KV-cache’. This acts like a short-term memory for the AI.

Instead of recalculating information, it reuses memory computed by earlier parts of the AI. This makes the process more efficient with less work.

Agentic Workflows and Licensing

Gemma 4 is also excellent at ‘agentic workflows’. This means the AI can perform actions, not just provide text.

It can use tools, write code, and even book flights or summarize news. Users can connect Gemma 4 to systems like OpenClaw to perform these tasks.

This capability is important because it offers an alternative to paid AI services. If a company restricts access to its AI, users can switch to Gemma 4. The model can be customized with specific instructions, often making it indistinguishable from paid options.

Expanded Context and Open Licensing

The context window for Gemma 4 has been increased to 256,000 tokens. This is twice as large as the previous version. While it can’t process entire movies, it’s enough to handle several long documents at once.

Perhaps most importantly, Gemma 4 uses the Apache 2.0 license. This is a very open license that allows for commercial use.

Users can modify, sell, and deploy the AI with very few restrictions. This is a significant change from the previous Gemma license, which had more limitations.

Limitations and Future

Despite its strengths, Gemma 4 has some limitations. It doesn’t have live access to databases, meaning it can sometimes provide incorrect information. It also struggles with highly complex, open-ended tasks and very detailed images.

However, the release of Gemma 4 is seen as a major benefit to the public. It’s a powerful AI tool that is freely available to everyone. Google DeepMind has stated that over 10 million downloads occurred in the first week alone.

This release comes at a time when access to some advanced AI models is becoming more restricted. Gemma 4 offers a powerful, open alternative for developers and users worldwide. Its accessibility and capabilities make it a significant development in the AI field.

Source: DeepMind’s New AI: A Gift To Humanity (YouTube)