Open source Mamba 3 arrives to surpass Transformer architecture with nearly 4% improved language modeling, reduced latency

This release is good for developers building long-context applications, real-time reasoning agents, or those seeking to reduce GPU costs in high-volume production environments.
IEEE

Vectorization of Narrow Matrix Multiplication for Ascend AI Inference Acceleration

Abstract: This research proposes and evaluates a novel approach to optimizing matrix multiplication (MatMul) on Huawei Ascend NPUs, motivated by a key insight: during matrix-vector multiplication ...