Atlas

Models

← All models

Tri-21B-Think Preview

Trillion Labs · 20.7B parameters

Tri-21B-Think Preview is an intermediate reasoning checkpoint of Tri-21B with 32K mid-training context expansion and instruction tuning for chain-of-thought reasoning and tool use. Its native context is 32,768 tokens, which the card says extends to 262,144 tokens with YaRN scaling.

Benchmark scores

BenchmarkScore
AA-LCR14.7
AA-LCR v1.121.7
AA-Omniscience Index-62.3
Artificial Analysis Omniscience Accuracy10.1
Artificial Analysis Omniscience Hallucination Rate80.5
Artificial Analysis Openness Index44.4
Artificial Analysis Progress Index v1197
Capability124.7
CritPt (avg@5)0.0
GPQA Diamond (Artificial Analysis avg@5)53.8
Humanity's Last Exam (Text-Only)5.8
IFBench (Artificial Analysis avg@5)47.1
SciCode (Artificial Analysis avg@3)17.8
Terminal-Bench Hard2.3
τ²-bench (Telecom)93.3
Loading Atlas data…