QUASAR: Lowering the Loss Floor of Quantization-Aware Training with Loss-Aware Reconstruction Paper • 2608.13966 • Published 22 days ago • 3
GLM-5.3-Flash · Alis MLX Collection Rewritten 2026-08-30: router/mHC/KDA aux kept unquantized. 4bit=QUASAR-init (KL -8.5% vs RTN). ~29-33 tok/s on M3 Ultra. Receipts in each repo. • 4 items • Updated 6 days ago
avlp12/Kimi-K2.7-Code-Alis-MLX-Dynamic-3.6bpw-VLM Image-Text-to-Text • 1T • Updated 8 days ago • 782 • 2
GLM-5.3-Flash · Alis MLX Collection Rewritten 2026-08-30: router/mHC/KDA aux kept unquantized. 4bit=QUASAR-init (KL -8.5% vs RTN). ~29-33 tok/s on M3 Ultra. Receipts in each repo. • 4 items • Updated 6 days ago