Minima Achieves 4-Bit Quantization for Gated DeltaNet Hybrid 27B LLM
September 2, 2026
Minima enables NVFP4 W4A4 quantization across all 496 layers of a 27B hybrid model, including the recurrent Gated DeltaNet blocks. The recipe matches BF16 perplexity across multiple benchmarks while reducing size to 17.5 GiB and increasing prefill speed by 14-19%.
HOW THIS AFFECTS YOU
●
builderYou can deploy highly compressed 27B hybrid models with minimal perplexity loss and faster prefill.
●
researcherThis demonstrates that recurrent state gates in hybrid architectures can survive aggressive 4-bit quantization.