#quantisationmethod

eicker.news ᳇ tech newstechnews@eicker.news
2025-10-05

#Huawei’s Computing Systems Lab introduced #SINQ, an #opensource #quantisationmethod for large language models (#LLMs). SINQ reduces #memoryusage by 60-70% without sacrificing output quality, enabling models to run on less powerful #hardware. The technique, available on GitHub and Hugging Face, uses #dualaxisscaling and #SinkhornKnopp-style #normalisation for improved performance. venturebeat.com/ai/huaweis-new #tech #media #news

Client Info

Server: https://mastodon.social
Version: 2025.07
Repository: https://github.com/cyevgeniy/lmst