BLAST RADIUS

Open-weight & owned stack · Quantization efficiency

llama.cpp Improves MUSA FlashAttention And Accelerator Support

Sep 25, 2026, first seen via Training & inference tooling releases

llama.cpp b11178 improved MUSA FlashAttention performance and enabled CUB-based sorting operations on Moore Threads MTT S5000 hardware.

Read the original at github.comOpens the publisher's site in a new tab

More in Open-weight & owned stack

Kyutai Releases Voice Of Reason CheckpointsSep 23Xiaomi Releases MiMo-V2.6 Open-Weight Model FamilySep 22Intrinsic open-sources Core robotics software infrastructureSep 22Harvey Launches Kimi-Based In-House ModelSep 21Kimi K3 Reaches AWS And Alibaba Cloud PlatformsSep 21