Memory Tensor releases and open-sources MemT 1.0 memory-native large model

Ayi AI Notes

MemT-1.0, an open-source native-memory large model, cuts 256K memory usage to a constant 40M

MemT-1.0 is open-sourced, connecting a native memory module in parallel within the Transformer to turn memory into a state in the forward computation flow, releasing up to 288B weights and a 66-page technical report. The authors claim that under a 256K ultra-long history, memory overhead drops from 5.06G to a constant 40M, KV cache is reduced by more than 120 times, and first-token latency stays stable at just over 100 milliseconds. The team also wrote more than a thousand code-correction experiences into the internal memory state, which do not occupy prompt tokens, and the pass rate on a standard code test set increased by 8% for three consecutive runs.

QbitAI

Memory Tensor Open-Sources MemT 1.0 Series Memory-Native Models

Memory Tensor has open-sourced and released the MemT 1.0 series of memory-native models, including MemT-1.0-Nano with 36B parameters and 3B activated (based on Qwen3.6-35B-A3B) and MemT-1.0-Flash with 288B parameters and 17B activated (based on DeepSeek-V4-Flash), adding a memory computation pathway parallel to attention in selected Transformer modules. The team built a 34-billion-token memory training corpus, of which 20 billion tokens were used for continued pre-training and 14 billion for post-training. In a comprehensive evaluation across 11 memory benchmarks, Flash and Nano achieved average scores of 50.6 and 47.0, respectively, ranking in the top two among the parametric memory methods evaluated; on 500 SWE-bench Verified tasks, the proportion of tasks that Nano succeeded on all three attempts rose from 54.00% for the no-experience baseline to 62.80%.

kr36_ai

English translation is being prepared.