Tri-21B-Think Preview: benchmark results
Preview of Trillion Labs' 21B RL-trained reasoning model with a backtracking mechanism (February 2026). Provider: Trillion Labs. Released 2026-02-10. Access: Open.
Aggregate Rating 1439 ± 41, rank #4121 of 10870 rated models, from 25 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AA TAU-2 Bench | 93.27 | Accuracy (%) | 90 |
| Tau2-Bench Telecom | 93.27 | Success Rate (%) | 89.8 |
| AA IFBench | 47.14 | Accuracy (%) | 55.3 |
| AA-Omniscience Hallucination Rate | 80.53 | Hallucination Rate (%) | 47.9 |
| AA Omniscience - Software Engineering (SWE) - Julia | 8 | Accuracy (%) | 46.1 |
| AA Omniscience - Software Engineering (SWE) - PHP | 18 | Accuracy (%) | 43 |
| AA Omniscience - Software Engineering (SWE) - Swift | 28 | Accuracy (%) | 40.1 |
| AA Humanity's Last Exam | 5.79 | Accuracy (%) | 33.3 |
| AA GPQA Diamond | 53.84 | Accuracy (%) | 27.8 |
| AA Long Context Reasoning | 21.67 | Accuracy (%) | 22.6 |
| AA Omniscience - Software Engineering (SWE) - Kotlin | 10 | Accuracy (%) | 22.4 |
| AA CritPt | 0 | Accuracy (%) | 20.7 |
Interactive version: theaggregate.ai/model?slug=tri-21b-think-preview · How It Works · Data refreshed daily, snapshot 2026-11-12.