Why Gated DeltaNet survives 4-bit: full NVFP4 W4A4 quantization of a hybrid 27B LLM

minima-ai · hf · 2026-09-04

minima-ai fully quantized a hybrid 27B LLM — including its recurrent Gated DeltaNet layers — to NVFP4 W4A4, preserving accuracy on long-context and reasoning benchmarks. The key is localizing outliers and exploiting the robust delta-rule dynamics, suggesting recurrent halves of hybrid LLMs are naturally low-bit friendly for edge deployment.

Original post →

More from Infra

Infra channel →