Gemini 4

Gemini 4 Argon's maximum output has been increased from 64,000 tokens to 1 million (referring to the output limit, not the context window)

Google published a new member of the Gemini series, Gemini 4 Argon, on its official blog.

The researcher believes that the value of general-purpose models like Gemini 4 lies in coordinating and orchestrating the work surrounding drug discovery, helping save time and resources, guiding researchers onto the right path, and ultimately accelerating better drugs reaching patients.

The researcher pointed out that Gemini 4 enabled him to transcend the limitations of his own area of expertise and uncover insights that would very likely have been overlooked if relying on a single person's reading alone.

The researcher estimated that generating reports for 10 different patents would require nearly a month of continuous work in the traditional way, whereas Gemini 4 completed the same task in a single afternoon.

The researcher gave Gemini 4 a set of patents he was very familiar with, and the model extracted key information into a spreadsheet and additionally generated a complete report, with output that was highly accurate and reached a directly usable level.

There is a division of labor between specialized models and general-purpose models in drug discovery: specialized models target specific tasks (such as structure prediction), while general-purpose large models (such as Gemini 4), with their common-sense reasoning and cross-domain connection capabilities, can serve as an orchestration layer integrating the outputs of specialized models with unstructured information such as literature and patents.

Gemini 4 not only summarized the content of the papers but also connected knowledge across different fields, providing insights beyond the papers themselves.

The researcher used Gemini 4 to summarize academic papers describing protein targets, and the model generated a high-quality report in about 10 to 15 minutes.

Gemini 4 can reason through video understanding, precisely locate the cause of failure, and translate corrective solutions into specific bodily instructions such as stance width, arm-swing trajectory, and release timing.

Traditional reinforcement learning requires engineers to carefully design reward signals, whereas Gemini 4's method bypasses this step and uses the results of video understanding as the learning signal.

The experiment demonstrated three key capabilities: error diagnosis driven by video understanding, cross-modal reasoning, and Sim-to-Real adaptation, forming a self-improvement loop without human intervention.

In the experiment, the Mojoku simulator served as a practice ground and data generator, producing a video record of each throw for Gemini 4 to analyze the cause of failure.

Gemini 4's Sim-to-Real transfer method performs online calibration through real-time visual feedback rather than relying on precise physical modeling.

In that medical software case, Gemini 4 identified a configuration file containing the application's master key, and through read-only operations turned the master key into administrator access, after which it assessed the blast radius.

In the Wiz research team's shadow A/B testing on real-world systems, Gemini 4 found more than twice as many vulnerabilities as other baseline models.

In long-duration security tasks, Gemini 4 can work continuously for hours, chaining together complex vulnerability signals and locating critical risks that could not be found using other SOTA models.

In one medical software case, Gemini 4 found a publicly accessible file download function that could be bypassed to download internal files from the underlying system.

The Wiz researchers concluded that, based on the cybersecurity capabilities demonstrated by Gemini 4, defenders now have the means to find and fix vulnerabilities faster and are able to stay ahead of attackers.

A benchmark chart circulating online purportedly showing Gemini 4 was confirmed to be AI-fabricated.