Disaggregated Quantization: Specializing LLM Prefill and Decode
Open paper
Benchmark
Model
Rank
Results
multimodal-reasoning-on-mmmu-pro
Qwen3.8-27B NVFP4 prefiller + Q3_K_XL decoder (extra data)
#18
Accuracy: 72.31
Rank counts only results with a code link.