MemT-1.0
MemTensor's memory-native models, a fixed-size memory state inside the model, on Qwen3.6 (Nano) and DeepSeek-V4-Flash (Flash).
In short. MemT-1.0 is a pair of open AI models with a built-in memory. Instead of re-reading the whole conversation each time, the model keeps a compact memory state that it updates after every turn and that you can save and load later. One model is built on Alibaba's Qwen3.6 and the other on DeepSeek-V4-Flash.
Summary
MemTensor, the company behind the MemOS memory framework, published MemT-1.0-Nano on Hugging Face on 15 October 2026. The Flash model followed on 17 October and the technical report on 19 October. Each model takes an existing mixture-of-experts model as its backbone, Qwen3.6-35B-A3B for Nano and DeepSeek-V4-Flash for Flash, and adds trainable memory parameters, 451M for Nano and 3.61B for Flash. A Local Memory Block holds a multi-channel memory matrix that changes as the conversation goes on, and a Hyper Memory Block adds trainable query, key and value projections with erase, write, forget and output gates. During generation, memory attention reads the current state and adds the result to the backbone's attention output, and at the end of each turn a Gated DeltaNet-2 update writes the new information into the state.
MemTensor says the persistent state stays the same size however long the history grows. At a 128K-token history, the report puts Nano at about 66 times less persistent storage than a full-context KV cache and about 1.86 times faster end to end. Across 11 memory benchmarks, the card says Flash and Nano rank first and second among the parametric-memory methods it evaluated, and the report calls them competitive with retrieval-augmented methods, while the card admits that strong text-based memory systems remain ahead overall, especially on long histories that need detailed evidence retrieval. A coding-agent demo stores retrieved experience in the native memory, so it takes no space in the task prompt.
Both models need custom code. The card asks for Python 3.11 or newer, a CUDA 13 build of PyTorch 2.11.0, Transformers 5.12.1 and trust_remote_code, and its sessions let you chat with memory enabled, save the state to a safetensors file and resume without the plaintext conversation. The weights are bf16 only, about 70.2 GB for Nano and 576 GB for Flash, and MemTensor publishes no quantized builds. Nano uses Apache 2.0 and Flash uses MIT, the same terms as their backbones. In the 30 days to 11 November 2026, Hugging Face counted about 89,000 downloads for Nano and 57,000 for Flash. The config files carry the backbones' position limits, 262,144 tokens for Nano and 1,048,576 for Flash, but the card states no context length of its own.
Good for
- Researchers studying memory that is built into the model instead of added through retrieval
- Assistants that must keep user facts across sessions without keeping the transcript in the prompt, using the save-and-resume memory file
- Long-running agents, since the card's coding setup reuses past experience without adding it to the task context
- Teams that want a memory model under Apache 2.0 or MIT on a known backbone
Watch out for
- The card says strong text-based memory systems still lead overall, especially when a task needs detailed evidence from a long history, so test against a retrieval baseline
- Memories are kept as latent state rather than readable text, and the card shows no way to inspect or delete a single memory
- It needs custom code, exact library versions and CUDA 13, and only bf16 weights exist, so the site's smaller 4-bit memory estimates do not apply
- The benchmark results are published as an image, with no numbers in the card text, and the speed and storage figures are MemTensor's own
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the MemT-1.0-Nano variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| MemT-1.0-Nano BF16 | 35.1B 3B active | 81 GB | 40 GB | 24 GB | Gaming GPU |
| MemT-1.0-Flash BF16 | 288B 17B active | 662 GB | 331 GB | 199 GB | Big-memory desktop |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
Frequently asked questions
- What is MemT-1.0?
- MemT-1.0 is a pair of open AI models with a built-in memory. Instead of re-reading the whole conversation each time, the model keeps a compact memory state that it updates after every turn and that you can save and load later. One model is built on Alibaba's Qwen3.6 and the other on DeepSeek-V4-Flash.
- Can I use MemT-1.0 commercially?
- Yes. MemT-1.0 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is MemT-1.0 open source?
- MemT-1.0 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run MemT-1.0 on my own computer?
- Yes, on a good graphics card. The smallest variant, MemT-1.0-Nano, needs about 24 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was MemT-1.0 released?
- MemT-1.0 was released by MemTensor on October 15, 2026.
- How long a context window does MemT-1.0 have?
- MemT-1.0 supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- MemT-1.0: Towards Memory-Native Evolution (arXiv, 19 Oct 2026) arxiv.org
- MemT-1.0-Nano model card (Hugging Face) huggingface.co
- MemT-1.0-Flash model card (Hugging Face) huggingface.co
- MemT-1.0 repository (GitHub) github.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Hy-Memory Tencent | 30.1B 3B active | 256K | Apache 2.0 (Permissive) | 20 GB Gaming GPU | |
| Qwen3.6 Alibaba Qwen | 27.8B – 36B 3B active | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| DeepSeek V4 DeepSeek | 284B – 1.6T 49B active | 1M | MIT (Permissive) | 176 GB Big-memory desktop | |
| Ornith 1.5 DeepReinforce | 9.65B – 403B 17B active | 256K | MIT (Permissive) | 6.66 GB Laptop or phone |