After China’s DeepSeek set the AI world alight with lightweight models that can provide pretty strong performance on relatively low-cost equipment like phones, Google wants developers to know it’s still in the game.
On Wednesday, the search giant unveiled the latest iteration in its Gemma series of open AI models, which it launched just over a year ago in a noteworthy shift away from a proprietary-only approach.
Gemma 3 comes in four sizes, ranging from a minuscule 1 billion parameters (the internal variables that determine how models function) to a still-svelte 27 billion parameters. According to Google, these are the company’s “most advanced, portable, and responsibly developed open models yet.”
What’s more, Google claims Gemma 3 outperforms DeepSeek-V3, Meta’s Llama-405B, and OpenAI’s o3-mini. This, it says, makes it the best model that can run on a single AI accelerator chip.
Gemma 3 also offers a much larger “context window” than its predecessors, meaning it can remember a lot more information at one time, allowing it to handle larger inputs. But at 128,000 tokens, that just means Google’s open models have caught up with rivals like Llama and DeepSeek.
Google also used the opportunity to release a Gemma 3–based image safety checker called ShieldGemma 2, which developers can use to identify things like sexual or violent content in pictures.
That wasn’t all. Google on Wednesday also revealed more about its push back into the world of robotics. Google’s DeepMind division has now wrangled its flagship Gemini 2.0 model into two new models dedicated to the field. (Parent company Alphabet took a step back a couple years ago when it called time on its Everyday Robots moonshot. In December, it announced a strategic partnership with a humanoid robotics firm called Apptronik.)
One, Gemini Robotics, can take natural-language instructions and turn them into actions, also taking into account changes to the robot’s environment. It is supposedly dexterous enough to handle tasks like folding origami and packing things into Ziploc bags. The other model, Gemini Robotics-ER, focuses on spatial reasoning, meaning it can do things like figure out the appropriate way to grasp and lift a coffee mug that is placed before it.
Both the robotics and Gemma 3 announcements come with liberal lashings of safety talk—and for good reason. Open models come with inherent safety risks, as they are not under the control of the company that releases them. Google said Gemma 3 had been thoroughly tested, with great emphasis on the models’ potential for making harmful substances, as the models supposedly have strong STEM skills. Meanwhile, robots can hurt people, so Gemini Robotics-ER is designed to judge the safety of the actions it enables and “generate appropriate responses.”