Early-Signal Detector

method · first seen 5 Oct 2026

NVFP4 quantization

Also: MX, NVFP4, NVIDIA FP4, 4-bit floating point, microscaling formats, FP4 microscaling formats, NVIDIA 4-bit floating point

A 4-bit floating-point number format with block scaling for quantizing neural network weights and activations on NVIDIA hardware.

NVFP4 is NVIDIA's 4-bit floating-point format (E2M1 values with a shared scale per 16-element block) that runs natively on Blackwell tensor cores. The sources show work moving past inference-only quantization toward training with it. Several groups apply it to RL post-training and report learner-sampler mismatch and the fixes they built for it. One comparative study, co-authored by Dan Alistarh, evaluates NVFP4 pretraining recipes at 8B and 30B MoE scale. Other papers cover smoothing and scale optimization, FP4 KV caches, LoRA fine-tuning on consumer GPUs, and per-tile format selection. Hobbyist repos ship vLLM and SGLang serving recipes for DGX Spark and RTX cards.

If NVFP4 training and RL prove stable at scale, 4-bit precision could become the default for both serving and training on current NVIDIA hardware, roughly doubling throughput over FP8. Caveats remain, such as slower acquisition of rare facts during FP4 pretraining.

Analyst note (emerging), 17 Nov 2026.

Mentions per week

0419 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 4 on 16 Nov 2026 · 8 mentions in 28 days · 8 independent groups · papers → code → discussion · first seen 6 weeks ago · acceleration +1.83

Earliest sources

Latest sources

Mentioned together with

quantizationreinforcement learning from policy gradients with train-inference mismatchGPTQKV cache quantizationreinforcement learning from policy rolloutsSGLangself-hosted model inferencespeculative decodingvLLMNVIDIA DGX Spark

In the ledger