MemT-1.0 brings native memory into the model, delivering 1.86x faster inference

The MemTensor team released MemT-1.0 (arXiv:2610.26825), an architecture that builds state updates directly into the model's compute graph instead of relying on external context storage. The MemT-1.0-Nano (36B total, 3B active) and MemT-1.0-Flash (288B total, 17B active) models were trained on 34B tokens; at 128K-token history, Nano delivers roughly 1.86x end-to-end inference speedup and a 66x reduction in persistent storage overhead.

• MemT-1.0-Nano: 36B total, 3B active parameters, 10M dynamic memory
• MemT-1.0-Flash: 288B total, 17B active parameters, 336M dynamic memory
• At 128K-token history: ~1.86x inference speedup and 66x lower storage overhead
• Native memory improved task completion and consistency on SWE-bench Verified and Multilingual