An AI Model from Nvidia Performs as Well as Closed Models in the Informatics Olympiad
44s
2
12:14, 21.10.2025
Nvidia announced that its AI model, built similarly to open-weight models, won gold at the International Olympiad in Informatics (IOI). Previously, only an experimental closed model from OpenAI had achieved such success.
New Strategy "GenCluster"
The basis of the success was the GPT-oss-120 B model with MoE (Mixture of Experts) architecture and the GenCluster strategy, which includes generation and clustering. This strategy allowed Nvidia to reallocate computational resources towards inference. The system first generated up to 5,000 solution variants for a problem; then, the solutions were grouped based on structure and behaviour; ultimately, the best solutions from each group entered a "competition," the winner of which was the final version submitted.
Economic Efficiency
Since the model uses the MoE architecture, the cost of a single run is significantly reduced. This allows Nvidia to perform thousands of generations more economically. Unfortunately, public data on the costs is not yet available.
Was this article helpful to you?
Yes
No
VPS popular offers
-10%
Linux
€ 7.7 /mo
Billed annually
-21.5%
Windows
€ 40 /mo
Billed annually
-10%
Linux
€ 25.25 /mo
Billed annually
-10%
Linux
€ 70.49 /mo
Billed annually
-10%
Windows
€ 10.23 /mo
Billed annually
-10%
Windows
€ 9.9 /mo
Billed annually
-10%
Linux
€ 8.99 /mo
Billed annually
-10%
Windows
€ 28.44 /mo
Billed annually
-9.5%
Windows
€ 121.5 /mo
Billed annually
-10%
Linux
€ 40.33 /mo
Billed annually
