Gemini Robotics 2.5 hits 96% at Mercedes-Benz

Gemini Robotics 2.5 builds on DeepMind's prior robotics models and was accelerated by Gemini 4 Argon, the frontier large language model rolled out in late September. DeepMind says Argon's capabilities let the robotics model learn new skills faster and apply them to real-world scenarios, and the new version was tested across over 200 distinct tasks.

The Mercedes-Benz proof-of-concept is notable because it moves the model from controlled lab conditions into production-adjacent work. Apptronik co-founder and CEO Jeff Cardenas said pairing Apollo 2 with Gemini Robotics let the models improve dramatically in just a few weeks of post-training, with very little drop-off when carried over to Mercedes-Benz production.

AlphaGenome 2 addresses a different problem: every cell carries the same DNA, but different genes are activated at different times. DeepMind used data from the Human Cell Atlas to give a view of how cells use genes, and the new model adds finer detail about specific cell types and states, plus DNA methylation. Professor Sarah Teichmann of the University of Cambridge said the model adapts directly to new single-cell data without retraining, even from very small samples.