TensorFold 1.3.0 boosts GLM-5.3-Flash to 104 tok/s, 24% faster on Blackwells

AIFlow_ML · x · 2026-11-20

TensorFold 1.3.0 is out: GLM-5.3-Flash jumps from 83.8 to 104.1 tok/s (about 24% faster) on two RTX PRO 6000 Blackwells at TP=2 with NVFP4 experts. On Metal, DeepSeek-V4.1's prompt-attention kernel is 1.5× faster. Flash Next frees 3 GiB of page cache on CUDA, Linux gains opt-in on-demand KV page mapping, and tensorfold models/info now estimate which checkpoints fit your machine accounting for context and parallel streams.

Original post →

More from Infra

Infra channel →