Preliminary internal results. Final benchmarks pending external review.
MANTLE builds explicit causal models, enabling robust counterfactual reasoning and planning under uncertainty.
Designed for long-horizon task decomposition. Maintains coherent goal structures across hundreds of reasoning steps.
Maintains performance when inputs differ significantly from training distribution — critical for real-world deployment.
Dynamically routes computation to specialized expert sub-networks, achieving ~120B param capacity at ~40B inference cost.
Produces calibrated confidence estimates, enabling downstream systems to route low-confidence outputs for human review.
256K token context with maintained coherence — critical for long-document reasoning and multi-session planning.
Terra-1 users automatically qualify for Mantle-1 early access at no extra cost. models@texasagilabs.com