Google has announced the addition of Gemma to its family of open source artificial intelligence models, and DeepMind has introduced Gemini Robotics, a model based on Gemini 2.0 designed for robotics.
Gemma 3
According to Google's website, Gemma 3 is a collection of lightweight open source models based on the same research and technology as Gemini 2.0. It is designed to run quickly on devices such as phones or laptops to help developers create AI applications. The new model is available in several sizes: 1B, 4B, 12B, and 27B.
Image: Google
Google claims that Gemma 3 outperformed Llama-405B, DeepSeek-V3, and o3-mini in preliminary human preference scores on the LMArena leaderboard, and can fit on a single GPU or TPU host. The new model offers:
- out-of-the-box support for more than 35 languages and pre-trained support for more than 140 languages;
- a context window with 128,000 tokens;
- support for function calls and structured output;
- creation of programs that analyze images, text, and short videos.
At the same time, Google is also releasing ShieldGemma 2, a 4B image security checker based on Gemma 3. This tool "provides a turnkey solution" for analyzing the safety of images by classifying them into three categories:
- dangerous content;
- explicit sexual content;
- violence.
The model can be downloaded via Kaggle and Hugging Face or Google Studio.
Gemini Robotics
DeepMind has demonstrated its progress in developing AI for robots. The researchers unveiled two new models based on Gemini 2.0, claiming that they "lay the foundation for a new generation of helpful robots."
The first model is Gemini Robotics, which facilitates direct control of robots. The company says that artificial intelligence systems for robots should have three characteristics:
- versatility - the ability to adapt to new situations (including those not foreseen during training);
- interactivity - the ability of a robot to respond to humans and the environment;
- dexterity - many tasks that humans perform automatically require precise motor skills that are difficult for robots to master.
Researchers have shown how robots can make origami, play tic-tac-toe, carefully fold glasses, and understand tasks that sound like natural language and perform them accurately despite obstacles.
Google is working with Apptronik, the developer of the Apollo bipedal robot, to create the next generation of humanoid robots.
The second model on display is the Gemini Robotics-ER, which has advanced spatial understanding. It allows roboticists to run their own programs using Gemini's reasoning capabilities.
