Tag
论文提出 FOCUS——一个基于 Mamba-2 的超轻量 EEG 认知负荷分类模型，采用固定 FFT 频带分解与权值共享编码，在仅 0.04M 参数下于 11 人试点实验中达到 74.75% 准确率与 0.08ms 前向延迟。
NVIDIA released the Nemotron 3.5 Super VL 120B-A12B model family — a 120B-total/12B-active vision-language model with Mamba-2/MoE hybrid architecture, C-RADIO vision encoder, 1M-token context, and native speculative decoding via MTP — available in BF16, NVFP4, and other variants under the OpenMDW-1.1 license.
A 4-bit quantization of NVIDIA's TwoTower block-diffusion LM (Nemotron-Labs-TwoTower-30B-A3B-Base) that compresses both towers into ~38 GB, enabling single-GPU inference at slow speeds (~2-4 tok/s).
This paper compares xLSTM, Mamba-2, and Gated DeltaNet on complex sequence modeling tasks and finds xLSTM superior due to its enhanced state tracking and memory dynamics, validated on synthetic length-generalization tasks.