MiniMax M2.5 Uses 10B Active Parameters per Token, Aiming for Cheaper Always-On Agents

MiniMax M2.5 hits about 80% on Sweetbench and runs near 100 tokens per second, helping teams deploy faster models on tighter budgets.
IEEE

ScanNow: A Scan Window-Based Sparse Matrix Multiplication Accelerator Design

Sparse matrix-matrix multiplication (SpMM) is a crucial kernel in various applications, including sparse deep neural networks [1]–[6], graph analytics [7], triangle counting [8], and linear algebra ...
IEEE

An Adaptive Sparse Matrix Compression CIM Accelerator based on 256Kb SOT-MRAM for Downlink Massive MIMO Communications

Abstract: Downlink precoding in massive multiple input multiple output (MIMO) systems involves high-dimensional sparse matrix calculations, which poses challenges to existing architectures.

MiniMax's new open M2.5 and M2.5 Lightning near state-of-the-art while costing 1/20th of Claude Opus 4.6

Chinese AI startup MiniMax, headquartered in Shanghai, has sent shockwaves through the AI industry today with the release of ...