88Z AI, 22 October — The Memory Tensor team has open-sourced the MemT 1.0 series of foundation models, including Flash and Nano versions. This series integrates updatable internal memory into computation, allowing access to historical information without retrieving original text. In 11 memory benchmarks, the two models achieved average scores of 50.6 and 47.0, ranking top two among parameterized memory methods. Technical highlights include Gated DeltaNet-2 per-token updates, MoS hybrid storage expansion, and a two-s
At a glance
- MemT 1.0 series is open-sourced, enabling internal memory access without retrieving original text
- Ranked top two among parameterized methods in 11 memory benchmarks, utilizing technologies like Gated DeltaNet-2
The update
The Memory Tensor team has officially released and open-sourced the MemT 1.0 series of memory-native foundation models, primarily consisting of the MemT-1.0-Flash and Nano versions. The innovation of this architecture lies in integrating continuously updatable internal memory directly into the computation process. This allows the model to access past information without retrieving original historical text, thereby enhancing efficiency and coherence. In terms of performance, the MemT 1.0 series demonstrated strong capabilities across 11 memory benchmarks. The two models achieved average scores of 50.6 and 47.0 respectively, ranking in the top two among all parameterized memory methods. Key technical highlights include the use of Gated DeltaNet-2 for per-token updates, capacity expansion through MoS hybrid storage mechanisms, and the application of a two-stage training strategy to optimize model performance.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 记忆张量MemTensorSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.