FP4 Quantization-Aware Training

A DeepSeek-V4 training step that prepares the model for very low precision by exposing it to quantization effects during training.

A model that was never trained to survive very low precision often breaks when it is compressed too far. FP4 quantization-aware training bakes that low-precision pressure into training so the deployed model loses less of itself when it finally runs at FP4-like constraints.

At a glance

Released

June 2026

Authors

DeepSeek-AI

Regime type

Optimization

Related modules

What It Is

This is a quantization-aware training step aimed at extremely low precision. Instead of treating quantization as a last-minute export trick, the model sees precision loss during training.

Why It Exists

It targets low-bit deployment quality, memory savings, and practical serving cost.

How It Works

The training loop injects low-precision behavior into the optimization path so the model learns weights and activations that remain more stable after compression.
FP4 Quantization-Aware Training flow
Request and weight flow
QAT teaches the model under the same precision pressure it will later face in deployment.
textdeploymentvaluepertrainingcostapproxfracDeltatextlow−bitqualityDeltatextQATcompute\\text{deployment value per training cost} \\approx \\frac{\\Delta \\text{low-bit quality}}{\\Delta \\text{QAT compute}}

Compared To Nearby Regimes

Unlike post-training quantization, the model is warned about future quantization damage during training instead of after the fact.

Limitations And Failure Modes

QAT adds training complexity and still may not save every capability if the target precision is aggressive enough.

Tags

References

  1. DeepSeek-AI. "DeepSeek-V4 Technical Report." 2026.