New Gemini 4 Argon Released. It Builds Realistic 3D Simulations and Analyzes Video Frame by Frame

The most notable update is the model's spatial reasoning. Instead of downloading or importing ready-made 3D files, Argon is capable of rendering three-dimensional objects and scenes purely through mathematical formulas and code.

In one test, the model was shown only one concept art of a dragon and asked to create its 3D model without using 3D editors. Argon wrote code in pure Python using mathematical Signed Distance Functions (SDF), which fully rendered the dragon in 3D space with correct proportions from all angles.

All our tests were done with a single prompt — the question was what the model could do "out of the box" and without hours of conversation with it.

First, we asked it to generate a black hole model. Gemini 3.1 in High mode produced a rather simple simulation.

And Gemini 4 in High mode produced a much more complex and detailed model, which accounted for many more parameters. Unfortunately, the initial model significantly slowed down the browser, so it had to be simplified, but even the simplified result is much more realistic than in Gemini 3.1.

Argon is also much better at creating 3D models for printing from photos. For example, we uploaded a photo of a table for it to generate a ready file for a 3D printer. Gemini 3.1 did a mediocre job — it offered a table without shelves and didn't consider the shape of the legs in the photo.

Павялічыць

Meanwhile, Argon immediately redrew the table with all shelves and depicted its actual shape. It also understood that the legs had an uneven shape and accounted for this when building the model.

Павялічыць

Argon gained the ability to analyze real video recordings frame by frame and transfer what it observed into the physical world. The models were shown a 20-second amateur video where a person throws a beanbag in the game of cornhole.

Argon captured the spatial coordinates of the movement, built a simulation, and transferred the code to the humanoid robot Apptronik Apollo, which then threw seven beanbags into the target consecutively in real life.

We don't have a robot, but we tested this with a video of a runner who first runs fast, then jogs. Gemini 3.1 built a simple simulation and, apparently, did not analyze the video frame by frame (judging by the description it gave in the chatbot).

Whereas Argon created a much more complex code that considers many more parameters and allows interactive adjustment of the runner.

• 8
• 1
• 0
• 1
• 6
• 0
• 0
• 0