by MemTensor
Open-weights memory foundation model from MemTensor with a fixed-size persistent memory state, built on Qwen3.6-35B-A3B.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
MemT-1.0 is a family of memory foundation models from MemTensor that build persistent, evolving memory into the model's computation. A Native Memory Block runs alongside the backbone attention. It combines a Local Memory Block (a dynamic multi-channel memory matrix) with a Hyper Memory Block (trainable query/key/value projections and erase, write, forget and output gates). The memory state is updated at the end of each turn or segment, so the model can use earlier history without re-reading the original context. According to the model card, the state stays constant in size and cuts storage by about 66× compared with a full-context KV cache at 128K history length.
The family has two models. MemT-1.0-Nano is built on Qwen3.6-35B-A3B (36B total, 3B active parameters, 451M trainable memory parameters). MemT-1.0-Flash is built on DeepSeek-V4-Flash (288B total, 17B active). On 11 memory benchmarks, the card ranks Flash first and Nano second among the evaluated parametric-memory methods. It also says strong textual-memory systems remain ahead overall, particularly on long histories that need detailed evidence retrieval. The project also demonstrates native memory for experience-guided agentic coding.
Nano is distributed on Hugging Face under Apache-2.0 and loads through Transformers with custom code. The card calls for Python 3.11+, a CUDA 13 build of PyTorch 2.11.0 and a pinned Transformers version. It is aimed at researchers and developers who want to experiment with parametric memory for long-running assistants and agents.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Open model that stores information across conversations in a compact internal memory state instead of rereading chat history. Comes in a Nano version (3B active) and a Flash version (288B total, 17B active). Strong on persistent memory benchmarks.
Closest neighbours by meaning, not just by category.
Model · MiniMax
Open-weights native multimodal MoE model from MiniMax, ~428B parameters (~23B active) with 1M-token context.
Model · Meituan
Meituan's open-weights 1.6T-parameter mixture-of-experts LLM (48B active), aimed at agentic coding and reasoning tasks.
Model · Alibaba
Alibaba's open-weights 2.4T-parameter MoE language model (95B active) from the Qwen3.8 family, with a hosted Max version.
Model · Meta
Meta's open-weights 30B dense model for local agents, tuned for tool use and long tasks, Apache 2.0 licensed.
Model · Meituan
Meituan's open-weights 1.6T-parameter multimodal MoE model (48B active) for agentic coding and visual tasks.
Model
Open model family (9B dense, 35B MoE, 397B MoE) trained with a self-improvement loop for reasoning, coding and agentic tasks.