#model compression

2 stories

#ai efficiency #alexander heinecke #apple mlx #bitcos #bonsai 2 27b #breaking #evangelos georganas #gpu optimization #inference speed #local inference #low-bit models #multimodal model

Bonsai 2 27B: 9x smaller, nearly full power
PrismML's new compressed model runs advanced AI on local devices without meaningful capability loss.
Prismml · 2m
New method shrinks AI models below the theoretical limit
Researchers break through a fundamental barrier in ternary AI compression by exploiting how weights actually distribute.
arXiv · 1m · 5 views