Model techniques map
Techniquesinference & servinginference quantization

specific method · filed under inference & serving

NVFP4 ModelOpt re-quantization

Re-quantizes selected linear layers, such as MLP or MoE expert linears, to NVFP4 while retaining other components at higher precision.

Also called NVFP4 ModelOpt re-quantization (W4A16), ModelOpt re-quantization, NVFP4 quantization of MoE expert linears, NVFP4 re-quantization.

sources
2
models
2
labs adopt it
2
strongest
optional

How sources treat it

One count per evidence span, weakest treatment to strongest.

optional 2

Documented in

GLM-5.3Qwen3.6-35B-A3B

Evidence

2 spans quoted from the sources, strongest treatment first.

only the MoE expert linears drop to NVFP4 while shared experts, attention, embeddings, and the early dense layers stay BF16.

optionalinference servingin GLM-5.3Inferact
zai-org/GLM-5.3 — 743B / 39B active · MOE · 1024K ctx·vLLM·2026-09-10

NVIDIA's ModelOpt re-quantization: MLP linears drop to NVFP4 (W4A16) while the attention linears and KV cache stay FP8

optionalinference servingin Qwen3.6-27B-NVFP4NVIDIA
Qwen/Qwen3.6-27B — 27B · DENSE · 256K ctx·vLLM·2026-07-02

Filed alongside

Other methods under inference & serving :: inference quantization.

FP8 KV-cache quantizationFP8FP4 KV-cache quantizationNVFP4 quantizationFour-Over-SixMXFP4 weight quantizationPost-training quantizationQuantizationBlock-scaled INT8 quantization with stochastic roundingFP8 E4M3 quantizationGGUFMax-based scalingMSE-based scalingMXFP8 activation quantizationNVFP4 KV-cache quantizationSSM cache quantizationAWQ INT4 weight quantization (W4A16)BF16 inferenceBlock-wise E4M3 FP8 weight quantizationChannel-wise quantizationDynamic activation scalingEmbedding and KV-cache quantizationEmpirical bits-per-element budget selectionFine-grained FP8 quantization