36B total / 3B active (Nano); 288B total / 17B active (Flash)
262,144 tokens (Nano); 1,048,576 tokens (Flash), plus persistent native memory across interactions
Apache-2.0 (Nano); MIT (Flash); original backbone and third-party license terms remain applicable
Family of memory-native text-generation models that makes persistent, dynamically evolving memory part of model computation rather than repeatedly supplying the original conversation history at inference. The release includes MemT-1.0-Nano, built on Qwen3.6-35B-A3B with 36B total and 3B activated parameters, and MemT-1.0-Flash, built on a further-post-trained DeepSeek-V4-Flash backbone with 288B total and 17B activated parameters. Nano adds 451M trainable memory parameters and maintains a 10M-element dynamic state, while Flash adds 3.61B trainable memory parameters and maintains a 336M-element state. A Native Memory Block runs alongside selected backbone-attention layers and combines a dynamic Local Memory Block with trainable Hyper Memory query, key, value, erase, write, forget, and output-gating projections. Stable Gated DeltaNet-2 performs token-level memory updates after each interaction, while Mixture-of-Storage provides a mechanism for scaling memory capacity. The released evaluation configurations use one memory expert per selected layer, producing persistent FP32 state sizes of approximately 40 MiB for Nano and 1,280 MiB for Flash regardless of accumulated history length. Training uses a two-stage continued-pretraining and post-training pipeline supported by 34B tokens. The session interface can write a turn into latent memory, clear its plaintext conversation context, answer later questions using the retained state, and save or restore that state; conventional plaintext-context chat is also supported. Across 11 memory benchmarks, Flash averages 50.60 and Nano 46.97, placing first and second among the evaluated parametric-memory methods and exceeding the strongest reported parametric baseline by approximately 12.5 and 8.9 points, although explicit textual-memory systems remain stronger overall on detailed long-history retrieval. At 128K tokens of history, Nano is reported to provide approximately 1.86× end-to-end inference speed and 66× lower persistent storage than a full-context KV cache; its measured speedup rises to approximately 3.99× at 256K. In experience-guided agentic coding, Nano reaches 69.33% average success on SWE-bench Verified when experience is stored before execution and 69.73% with harness-triggered storage, compared with 66.07% without experience. On SWE-bench Multilingual, the corresponding results are 62.89% and 63.56% versus 60.00% without experience. The BF16 repositories contain approximately 70.2 GB of safetensors for Nano and 576.0 GB for Flash and require custom Transformers code. Nano natively supports 262,144 tokens and includes a packaged chat template; Flash supports 1,048,576 tokens and provides the custom MemT session/chat interface.