News › 2026 › April 2026
FP4 training shown stable at scale
Research2026-04-30
Source: arxiv.org
Researchers demonstrate stable four-bit training on Blackwell, cutting memory and energy for large runs without hurting quality.
Lower precision means more model per GPU. This work shows four-bit floating point can train large models to the same quality as higher precision, given the right scaling and loss handling.
On Blackwell hardware that translates directly into cheaper training; the parts are on the prices page and rentable via hosting.