Google releases Gemini 4 Argon model, dramatically reducing quantum computing resource requirements and boosting robot performance

Google recently launched Gemini 4 Argon, a model that claims to reduce resource requirements in quantum computing operations by more than 100 times, while helping researchers develop new mathematical solutions for fusion problems and improving real-world robot performance. In a demonstration, Argon generated a quantum algorithm component with a spacetime cost more than 100 times lower than previously published methods. The company said the implementation was formally verified by the Lean system, a system for checking mathematical proofs.

Robot Performance Improvements

The model also helped researchers find new solutions to a decades-old mathematical problem related to stellarator fusion reactors. In separate robot tests, training data generated by Argon simulation raised the robot success rate in real-world tasks from 56% to 75%. Google announced these advances on October 13 as part of the Gemini 4 Argon launch.

In another test, the model ran unattended on Google's existing software codebase for 11 hours, generating changes that reduced session preparation overhead by about 40% and increased package import speed by 21%. Quantum computing becomes more streamlined; quantum computers require carefully designed operations to perform calculations while limiting errors and hardware requirements. Some algorithms rely on smaller mathematical components called subroutines, which consume large amounts of computational resources. Gemini 4 Argon is now available to developers and enterprises worldwide on the paid tiers of the Gemini API and the Gemini platform, to Google Antigravity Ultra subscribers, and

Ultra and Pro subscribers of the Gemini app and Search can also use it.

Quantum Computing Resource Optimization

Google said Argon optimized the arcsin subroutine used in quantum simulation algorithms. Compared to the published quantum CORDIC implementation, the new version reduced spacetime cost by more than 100 times. Spacetime cost measures the computational resources required to execute an operation, accounting for hardware requirements and execution time. This improvement applies to the specific subroutine tested, not the overall quantum computing system. Argon also optimized an internal version of Google's machine-learning-based quantum error correction decoder, AlphaQubit.

According to Google, by converting it into a single hardware-optimized kernel, the model nearly doubled inference throughput.

The model's software engineering capabilities were tested on Google's open-source Agent Development Kit, a framework for building and deploying autonomous agents. In an 11-hour overnight run, Argon examined a codebase containing approximately 205,000 lines of code across nearly 800 files. It proposed six optimizations, two of which were accepted and merged into the project after human review. These changes reduced in-memory session preparation overhead by about 40% in long-running sessions and improved package import speed by 21%.

In an independent benchmark, the execution time of a 158-turn agent dropped from 65.6 seconds to 56.1 seconds. However, the time from launching a new process to its first model request remained essentially unchanged.

On fusion research, Google collaborated with University of Maryland professor Matt Landreman to identify a broader set of counterexamples refuting Grad's conjecture, a mathematical hypothesis dating to 1967. Earlier counterexamples had already challenged the conjecture. The newly identified solutions could provide reference cases for testing simulation software used to design stellarator fusion reactors, which use complex magnetic fields to confine hot plasma. Separately, Argon demonstrated the ability to generate physical robot training environments.

Starting from a single photo of a tabletop, the model built a physics-based simulation and generated more than 4,500 robot manipulation task training runs.

Gemini 4 Argon enabled a robot to learn to play cornhole just by watching video. It mapped the physics, coded its own simulation, and taught Apptronik's Apollo humanoid robot to land 7 consecutive successful throws. Google reported that the generated training data raised real-world robot success rates from 56% to 75%, while reducing task completion time from 35 seconds to 24 seconds. In another demonstration, Argon helped Boston Dynamics' Spot robot practice opening a pedal bin in simulation before attempting it with the physical object.

Beyond robotics research, researchers at Gamow Labs also used Argon to examine de-identified genomic data from newborns with a rare developmental lung disease. Google said the model identified a chromosomal rearrangement that standard software had failed to detect for years. These results span scientific research, software engineering, and physical robot control. However, the reported improvements remain specific to the experiments described, and broader performance in these areas has not been established.