AI model startup Trillion Labs said it has unveiled the Think model Tri 21B Think, which maximizes reasoning capabilities by applying Reinforcement Learning to the existing Tri 21B, and that it was listed within the global top 30 on the leaderboard of Artificial Analysis (AA), a global AI model performance analysis organization.
Tri 21B Think unfolds reasoning steps as tokens in the problem-solving process and, when needed, implements a backtracking structure that returns to a previous step for review on its own. The backtracking structure operates based on test-time scaling technology, meaning the longer the AI thinks, the better it gets at solving problems.
Through this, Tri 21B Think can deliver high performance in areas that surpass the limits of existing large language models (LLMs), including complex data analysis (deep research), autonomous AI agents, and advanced mathematics and coding, the company said.
Tri 21B Think has 2.1 billion (21B) parameters and is designed to fit on a single graphics processing unit (GPU).
Shin Jae-min, CEO of Trillion Labs, said, "We are very pleased that this AA listing allows Korea to contribute to the process of leaping into the global top three in AI," and added, "Having verified the technology that can deliver overwhelming performance even with limited resources, we will aim for the world's No. 1 performance, increase the model's completeness, and solidify our position as a global AI model beyond a national representative."