MemTensor open-sources MemT 1.0 Flash and Nano, models with built-in trainable memory state
MemTensor has open-sourced MemT 1.0 Flash and Nano, models with a built-in trainable memory state that keeps long histories in a fixed state. The company reports about 4x faster inference at 256K history and 130x smaller storage.
MemTensor has open-sourced MemT 1.0 Flash and Nano, two models with a built-in trainable memory state. The company reports about 4x faster inference at 256K history and 130x smaller storage, as covered by Pandaily.
What was released
MemT 1.0 comes in two variants, Flash and Nano. MemTensor describes them as memory-native models: they keep long histories in a fixed state rather than carrying the full history forward.
Reported figures
All figures below are MemTensor's own claims.
- Inference speed: about 4x faster at 256K history
- Storage: 130x smaller
Parameter counts, pricing, benchmark scores and training cutoff are not stated in the available summary.
What this means
The reported gains target two costs of working with long histories: inference speed and storage. A trainable memory state that stays fixed in size is the mechanism MemTensor points to for both. The 4x and 130x figures are company claims and have not been independently verified.
Related Articles
China Telecom open-sources Xing4.0-29B-Intent, a 29B intent model, with 2.753M-sample training corpus
China Telecom's Xingchen General AI Lab has released and open-sourced Xing4.0-29B-Intent, a model that turns a user's sentence into structured JSON covering intent and parameters. The release includes model weights, a 2.753 million-sample training corpus and an inference image. In the company's own tests, it scores 58 against a top score of 49 among general-purpose models.
Ai2 releases OlmoEarth v1.3 satellite embedding model with open global 10m embeddings
Ai2 released OlmoEarth v1.3, an open satellite embedding model that turns a year of Sentinel-1, Sentinel-2 and Landsat data into a 128-dimensional vector per 10-meter pixel. Ai2 claims it scores highest on 20 of 27 evaluations against AlphaEarth Foundations and Tessera v2, and it trained on 53% of the compute used for v1.2.
Hugging Face releases Carbon-X, a 14B-parameter open model that answers natural-language questions about DNA
Hugging Face released Carbon-X, a 14-billion-parameter open model that takes DNA sequences and English questions and returns conversational answers. It was trained on 300 billion tokens and is designed to run locally. Its developers say it has not yet been validated with wet lab experiments.
Ant Digital Technologies releases LingDT-Decision-1.0 in 4B and 27B sizes for direct structured decisions
Ant Digital Technologies has released LingDT-Decision-1.0, a model that makes structured choices directly instead of generating text. It comes in 4B and 27B sizes, and the company says it cuts token use by more than 90%.
Comments
Loading...