Curated LLM quantization papers covering GPTQ, AWQ, GGUF, NF4, FP4, NVFP4, MXFP4, and what low-bit inference means for GPU choice.
Quantization is the most direct way research turns into GPU buying decisions. A model that is impossible at FP16 may fit comfortably at Q4, while a poorly supported low-bit format can save VRAM but fail to improve real tokens per second.
This collection starts with practical local formats such as llama.cpp and AWQ, then moves into FP4, NVFP4, MXFP4, NF4, and layer-sensitive methods. The goal is to help you separate model file size from deployable inference speed.
Read these first if you want the fastest path from research to a hardware decision.
Uygar Kurt
A unified empirical evaluation of llama.cpp quantization formats for Llama-3.1-8B-Instruct on commodity hardware.
GPU Hunter takeaway: A GPU recommendation is incomplete without naming the quantization formats that are realistic for that VRAM tier.
Ji Lin, Jiaming Tang, Haotian Tang, Song Han
A hardware-friendly weight-only quantization method that protects salient channels based on activation statistics.
GPU Hunter takeaway: A practical paper for understanding why some 4-bit models are much faster than others on the same GPU.
Elias Frantar, Saleh Ashkboos, Torsten Hoefler, Dan Alistarh
A one-shot weight quantization method that uses approximate second-order information to compress large GPT-style models to 3-4 bits.
GPU Hunter takeaway: A 24GB card becomes far more useful when the model can be loaded at 4-bit without a major quality collapse.
Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, Luke Zettlemoyer
A finetuning recipe that backpropagates through a frozen 4-bit model into LoRA adapters using NF4 and paged optimizers.
GPU Hunter takeaway: If a GPU can barely fit a base model for inference, QLoRA explains how adapter training can still be possible.
These papers go deeper into kernels, cache policy, low-bit formats, or serving tradeoffs.
Musa Cim, Burak Topcu, Mahmut Taylan Kandemir
A component-wise sensitivity study of NVFP4 and MXFP4 quantization across Qwen2.5 model scales and transformer blocks.
GPU Hunter takeaway: Treat FP4 as a hardware capability that still needs model-aware quantization policy, not a universal speed switch.
Vage Egiazarian, Roberto L. Castro, Denis Kuznedelev, Andrei Panferov, Eldar Kurtic, Shubhra Pandit
A study of MXFP4 and NVFP4 post-training quantization that introduces Micro-Rotated-GPTQ and reports GPU kernel-backed speedups.
GPU Hunter takeaway: RTX 5090 and Blackwell FP4 claims should be judged against real FP4 kernels and accuracy, not just advertised tensor formats.
Saleh Ashkboos, Amirkeivan Mohtashami, Torsten Hoefler, James Hensman
A rotation-based quantization method that removes hidden-state outliers and quantizes weights, activations, and KV cache.
GPU Hunter takeaway: Good background for why next-gen GPUs may make activation and KV-cache precision just as important as weight precision.
The complete paper set for this topic, with source links and GPU Hunter takeaways.
Musa Cim, Burak Topcu, Mahmut Taylan Kandemir
A component-wise sensitivity study of NVFP4 and MXFP4 quantization across Qwen2.5 model scales and transformer blocks.
GPU Hunter takeaway: Treat FP4 as a hardware capability that still needs model-aware quantization policy, not a universal speed switch.
Vage Egiazarian, Roberto L. Castro, Denis Kuznedelev, Andrei Panferov, Eldar Kurtic, Shubhra Pandit
A study of MXFP4 and NVFP4 post-training quantization that introduces Micro-Rotated-GPTQ and reports GPU kernel-backed speedups.
GPU Hunter takeaway: RTX 5090 and Blackwell FP4 claims should be judged against real FP4 kernels and accuracy, not just advertised tensor formats.
Uygar Kurt
A unified empirical evaluation of llama.cpp quantization formats for Llama-3.1-8B-Instruct on commodity hardware.
GPU Hunter takeaway: A GPU recommendation is incomplete without naming the quantization formats that are realistic for that VRAM tier.
Ji Lin, Jiaming Tang, Haotian Tang, Song Han
A hardware-friendly weight-only quantization method that protects salient channels based on activation statistics.
GPU Hunter takeaway: A practical paper for understanding why some 4-bit models are much faster than others on the same GPU.
Elias Frantar, Saleh Ashkboos, Torsten Hoefler, Dan Alistarh
A one-shot weight quantization method that uses approximate second-order information to compress large GPT-style models to 3-4 bits.
GPU Hunter takeaway: A 24GB card becomes far more useful when the model can be loaded at 4-bit without a major quality collapse.
Guangxuan Xiao, Ji Lin, Mickael Seznec, Song Han
A post-training quantization approach that moves activation outlier difficulty into weights so W8A8 inference can stay accurate.
GPU Hunter takeaway: Read this when comparing INT8, W8A8, and FP8 serving paths on NVIDIA workstation or datacenter GPUs.
Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, Luke Zettlemoyer
A finetuning recipe that backpropagates through a frozen 4-bit model into LoRA adapters using NF4 and paged optimizers.
GPU Hunter takeaway: If a GPU can barely fit a base model for inference, QLoRA explains how adapter training can still be possible.
Saleh Ashkboos, Amirkeivan Mohtashami, Torsten Hoefler, James Hensman
A rotation-based quantization method that removes hidden-state outliers and quantizes weights, activations, and KV cache.
GPU Hunter takeaway: Good background for why next-gen GPUs may make activation and KV-cache precision just as important as weight precision.
Vage Egiazarian, Andrei Panferov, Denis Kuznedelev, Dan Alistarh
AQLM applies learned additive quantization to push LLM compression into the 2-3 bit range with practical CPU and GPU implementations.
GPU Hunter takeaway: Useful for builders choosing between a smaller dense model and an aggressively compressed larger model.
Start with the llama.cpp quantization evaluation, AWQ, GPTQ, and QLoRA. Then read the FP4 and NF4 papers when hardware support or Blackwell-class GPUs matter.
No. Quantization reduces memory footprint, but speed depends on dequantization, packed-weight kernels, memory bandwidth, and whether the runtime supports that format well.