Metis-1.6 Uses Hybrid Mamba-2 and MoE Architecture
October 26, 2026
Metis-1.6 features two models, Praxis (2.94B) and Logos (10.23B), utilizing a hybrid Mamba-2 and attention mixture-of-experts stack. Each model processes every token twice across 1.02T tokens of training data to optimize efficiency over parameter count.
HOW THIS AFFECTS YOU
●
builderYou can investigate these architectures for high-performance, low-cost model deployment.
●
researcherThe hybrid Mamba-2 and MoE architecture offers a new way to evaluate efficiency-to-performance ratios.