Google’s Gemini AI has made significant strides in the past year, allowing developers to play around with its new features. The latest version, Gemma 4, is here, and it brings some exciting upgrades. Unlike its predecessor, Gemma 3, which has started to feel outdated, Gemma 4 comes in four sizes designed for local use, making it accessible for many developers.
This new model features two large variants: the 26B Mixture of Experts and the 31B Dense. These can run on a powerful 80GB Nvidia H100 GPU. While that GPU costs around $20,000, it allows developers to use AI directly on their machines. With adjustments for lower precision, these models could even work on regular consumer GPUs. This means more people can tap into the potential of AI without needing expensive hardware.
One of the key improvements in Gemma 4 is reduced latency. The 26B Mixture of Experts activates just 3.8 billion of its 26 billion parameters during processing, which boosts its performance. The 31B Dense model focuses on delivering high-quality output but may be slower. Google encourages developers to customize these models for specific applications, allowing for greater versatility.
The other two models, Effective 2B (E2B) and Effective 4B (E4B), target mobile devices. They were designed to keep memory usage low, making them ideal for smartphones and smaller computers like Raspberry Pi and Jetson Nano. Working closely with Qualcomm and MediaTek, Google minimized battery usage and claims to have achieved “near-zero latency.” This is a notable upgrade over Gemma 3, making AI tools more efficient on mobile devices.
These advancements put Gemma 4 ahead of its predecessor. Google asserts it’s one of the most capable AI models for local hardware. In fact, the 31B Dense model is set to rank third on the Arena list of top open AI models, following GLM-5 and Kimi 2.5. It’s worth noting that even the largest Gemma 4 model is much smaller than those competitors, making it more affordable to run.
As AI continues to evolve, Google is addressing developer concerns about the licensing issues that have historically frustrated many in the tech community. By eliminating the custom Gemma license, Google is fostering a more open environment for innovation. This shift could encourage more developers to experiment and create new applications using AI.
In a recent survey by arstechnica.com, 67% of developers expressed a desire for more clear and flexible licensing models in AI technology. With the introduction of Gemma 4, Google seems to be responding to this demand, paving the way for more widespread use of AI tools.
