Korean AI startup TrillionLabs has entered the global top 30 on Artificial Analysis leaderboard with...
The AMW Read
TrillionLabs updates the player map with a high-performing reasoning model that exemplifies the structural shift from scaling parameters to scaling inference-time compute (test-time scaling).
Korean AI startup TrillionLabs has entered the global top 30 on Artificial Analysis leaderboard with Tri 21B Think, a 21-billion-parameter model that uses reinforcement learning and Test-time Scaling to enable backtracking during reasoning. The model can now pause, revisit earlier steps, and correct its reasoning path - a capability that improves performance exponentially as computation time increases, supporting complex data analysis, autonomous agents, and advanced math/coding tasks. This demonstrates how the AI industry is shifting from simply scaling model size to scaling inference-time compute, allowing smaller, more efficient models to compete with giants through deeper reasoning. Korea's push for AI sovereignty gains momentum as TrillionLabs proves it can achieve world-class performance with limited resources through innovative training methods like XLDA and rBridge.
#AI #TrillionLabs #TestTimeScaling #ReasoningModels #KoreaAI #ArtificialAnalysis