AI model startup TrillionLabs said on Feb. 19 it unveiled Tri 21B Think, a Think model that applies reinforcement learning to its existing Tri 21B to strengthen reasoning capabilities.

The company said Tri 21B Think was listed in the global top 30 on the leaderboard of Artificial Analysis, a global AI model performance evaluation group.

Tri 21B Think is based on TrillionLabs' existing Tri 21B model and uses reinforcement learning to advance its Thinking capability, which expands the reasoning process. It goes beyond generating an answer by unfolding reasoning steps in token form during problem-solving and implements a backtracking structure that returns to earlier steps for review when needed.

The backtracking structure operates based on Test-time Scaling technology, and the longer the AI thinks, the ability to solve complex and difficult tasks improves exponentially. With this, Tri 21B Think supports complex data analysis (deep research), autonomous AI agents, and advanced mathematics and coding.

Jae-min Shin (신재민), CEO of TrillionLabs, said, "I am very pleased that this AA listing enables us to contribute to Korea's process of leaping into the global top three in AI," and added, "Having confirmed the technological capabilities to deliver overwhelming performance with limited resources, we will raise the model's level of completion with the goal of No. 1 global performance and solidify our position as a global AI model beyond a national 대표-level model."