Metis launches open-source MemT 1.0 memory-native models

Metis has released and open-sourced the MemT 1.0 series of memory-native foundation models with persistent, updatable internal memory. MemT-1.0-Flash (288B total, 17B active parameters) and MemT-1.0-Nano (36B total, 3B active) scored 50.6 and 47.0 on average across 11 memory benchmarks, ranking first and second among parametric memory methods.

• MemT-1.0-Flash: 288B total, 17B active parameters
• MemT-1.0-Nano: 36B total, 3B active parameters
• Average scores of 50.6 and 47.0 on 11 memory benchmarks
• Trained on a 34B-token corpus