LeapQuant: Training-Free Quantization Speeds Up Linear Attention

LeapQuant is the first training-free method to quantize linear attention's recurrent state to 8/4-bit with negligible accuracy loss, delivering up to 4.25x kernel speedup and doubled throughput on hardware like the B200.

2026-11-12 ~ 2026-11-13 · 4 related posts

• LeapQuant: training-free quantization makes linear attention up to 4.25x faster — tianyin_xu · 2026-11-12
• LeapQuant: training-free 4/8-bit linear attention quantization, 4.25x faster kernels — berkeley_ai · 2026-11-13
• LeapQuant: 8/4-bit linear attention states, training-free, 2x throughput — SonglinYang4 · 2026-11-13

1 near-duplicate retellings: zealcaiden