Open Models Almanac
Methodsoptimizationoptimizer

single mention · optimization

Hybrid Newton–Schulz iterations

A hybrid Newton–Schulz scheme approximately orthogonalizes a matrix by running the iterations in two stages with different coefficients.1,2 The first stage uses coefficients that drive rapid convergence toward singular values near 1, and the final stage switches to coefficients that stabilize the singular values precisely at 1.3

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. The Newton-Schulz iterations aim to approximately orthogonalize — DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
  2. Our hybrid Newton-Schulz performs 10 iterations over two distinct stages — DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
  3. During the first 8 steps, we use coefficients to drive rapid convergence, bringing the singular values close to 1. In the final 2 steps, we switch to coefficients , which stabilize the singular values precisely at 1 — DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence

Filed under optimization › optimizer