Tri-21B-Think: benchmark results
Trillion Labs' open 21B Korean reasoning model (February 2026), the full release with backtracking chain-of-thought that runs on a single GPU. Provider: Trillion Labs. Released 2026-02-10. Access: Open.
Aggregate Rating 1497 ± 39, rank #2402 of 10916 rated models, from 22 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Tau2-Bench Telecom | 80.99 | Success Rate (%) | 72.2 |
| AA TAU-2 Bench | 80.99 | Accuracy (%) | 71.5 |
| AA IFBench | 54.63 | Accuracy (%) | 65.4 |
| AA Omniscience - Software Engineering (SWE) - HTML | 30 | Accuracy (%) | 51.7 |
| AA CritPt | 0.29 | Accuracy (%) | 45.9 |
| AA Omniscience - Software Engineering (SWE) - Kotlin | 14 | Accuracy (%) | 39.8 |
| AA Omniscience - Software Engineering (SWE) - R | 6 | Accuracy (%) | 36.2 |
| AA Omniscience - Software Engineering (SWE) - TypeScript | 13.33 | Accuracy (%) | 36.2 |
| AA GPQA Diamond | 60.1 | Accuracy (%) | 34.8 |
| AA Humanity's Last Exam | 5.93 | Accuracy (%) | 34.3 |
| AA Omniscience - Software Engineering (SWE) - Java | 12 | Accuracy (%) | 33.1 |
| AA Omniscience - Software Engineering (SWE) - Julia | 4 | Accuracy (%) | 30.4 |
Interactive version: theaggregate.ai/model?slug=tri-21b-think · How It Works · Data refreshed daily, snapshot 2026-11-08.