Early-Signal Detector

method · first seen 6 Oct 2026

FP4 quantization

Also: 4-bit floating point quantization

Representing weights or activations in 4-bit floating point format to reduce compute and memory cost.

FP4 quantization stores and computes weights, activations, KV caches or gradients in 4-bit floating point, typically with block-wise scale codes as in MXFP4 and NVFP4, a format that Blackwell-class GPUs execute natively in tensor cores. The new work in these sources moves beyond plain weight compression. Two papers, one with Qwen-affiliated authors, target FP4 rollouts for reinforcement learning on mixture-of-experts models. Others cover FP4 communication compression for model-parallel training, residue and low-rank corrections that recover accuracy (ResServe, BiLoCo), second-order scale search, and FP4 KV-cache attention for multi-head latent attention, including an NVIDIA paper. One open-source repo implements FP4 KV-cache decoding for vLLM.

Native FP4 hardware support turns accuracy-recovery methods into direct savings in serving and training cost. If these recipes hold up for RL post-training and attention kernels, FP4 could become the default precision for much of the LLM pipeline rather than an inference-only option.

Analyst note (emerging), 17 Nov 2026.

Mentions per week

0319 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 32 on 16 Nov 2026 · 8 mentions in 28 days · 8 independent groups · papers → code · first seen 6 weeks ago · acceleration +0.44

Earliest sources

Latest sources

Mentioned together with

mixture of expertsquantization-aware trainingpost-training quantizationlow-rank correctionSVDQuantsparse attentionvideo diffusion transformerFlashAttentionHadamard transformcommunication compression for model-parallel training