This work challenges the assumption that recurrent linear-attention layers require higher precision than other model components, showing that the Gated DeltaNet (GDN) blocks in a hybrid 27B-parameter LLM can be fully quantized to 4-bit weights and activations without accuracy loss. The researchers identify a four-part mechanism that protects GDN under quantization, including block-level scaling that confines outliers and delta-rule recurrence that actively erases state error. The quantized model matches BF16 performance on reasoning and retrieval tasks while cutting model size by 2.9x and improving prefill speed by 14-19%. The study also surfaces serving-stack mismatches that affect hybrid-model deployments, inverting some conventional quantization assumptions for this architecture.
