1 of 19

Fantastic Pretraining Optimizers

and Where to Find Them

A Fair Comparison of Modern Optimizers + Hyperball

Wen, Hall, Ma, Liang (2025) | Wen, Dang, Lyu, Ma, Liang (2026)

2 of 19

This Talk

Finding 1

With fair tuning across multiple scales, matrix-based optimizers including Muon & Soap yield clear speedup over AdamW.

However, all modern optimizers reach only 10% speedup over AdamW at 1.2B scale / ~100B tokens mysteriously.

Finding 2

The disappearing speedup is (partly) due to a constant weight decay (~0.1) shaping the effective learning rate.

Hyperball (weight-normalized training) eliminates this effect — persistent 20-30% speedup across scales.

Fantastic Optimizers and Where to Find Them (2025)�Fantastic Optimizers and Where to Find Them 2.1 (2026)

3 of 19

Pretraining as an Optimization Task

Pretraining is the mostly costly component of LLM training

Many recent optimizers claim 1.4-2x speedup over AdamW: Sophia [Liu et al.’24], SOAP [Vyas et al.’25], Muon [Liu et al.’25], MARS [Yuan et al.’25], C-AdamW / C-Lion (Cautious Optimizers) [Liang et al.’25], FOCUS [Liu et al.’25], SCION [Pethick et al.’25], SWAN [Ma et al.’25]

​

​

Puzzling phenomenon:

Why is AdamW still popular despite these claims?

A 1.5x speedup could save ~$20M on LLaMA 3.1 405B training

​

​

4 of 19

The Optimizer Landscape

Baseline: Adam(W)

  • First & second moment estimates
  • Default optimizer for years

Scalar Variants

  • Nesterov Adam (NAdam)
  • MARS, Cautious
  • scalar-preconditioned: each scalar parameter only uses its own gradient information

AdamW

NAdam

5 of 19

The Optimizer Landscape

Matrix-Based

  • Shampoo, SOAP, Muon, Kron
  • Matrix-level gradient info
  • Higher per-step cost

Shampoo

Scalar-Based

  • AdamW
  • Nesterov Adam (NAdam)
  • MARS, Cautious, Lion

Change Subject

Shampoo

Muon

6 of 19

Confusing Results in the Literature

Different evaluation protocols lead to vastly different conclusions about optimizer effectiveness.

Muon can't be both worse than AdamW AND 2x better!

Figure 1 of MARS

Figure 1 of Moonlight

Key Reason: Optimizers are not evaluated under respectively best practices!

7 of 19

Methodology: Three-Phase Fair Tuning

Fairness Principles

  1. Tune each optimizer separately
  2. Compare at end-of-training
  3. Evaluate across model scales and data-to-model ratios

Phase I

Fine-grained sweeps for every hyperparameter per-optimizer

Phase II

Sweeping�scale-sensitive hypers�for larger models or more data

Phase III

Hyper scaling laws�extrapolating to 1.2B�models

8 of 19

Phase I: Coordinate Descent Algorithm

Algorithm:

Coordinate descent on a discrete hyperparameter grid: tune one hyperparameter at a time, keep updates only if validation loss improves and stop when no single change helps.

Settings:

0.1B model 1-8x Chinchilla

0.3B, 0.5B 1x Chinchilla

1 Chinchilla = 20x token per parameter

Results: One table on the sensitivity of all hyperparameters for each settings

​

​

9 of 19

Phase I: Coordinate Descent Results

Observations:

  1. Loss is highly sensitive to hypers like learning rate.
  2. Different optimizers prefers very different hyperparameter.
  3. Loss is sensitive only to a subset of hyperparameters; many have negligible effect near optimum. Among sensitive ones, most have constant optimal values across scales.

Takeaway:

Narrow later sweeps to only scaling-sensitive hyperparameters like learning rate.

​

​

10 of 19

Phase II: Sweeps on 0.1B - 0.5B & 1-8x Chinchilla

​

Speedup:

Additional percentage of tokens AdamW needs to match the performance of optimizer X

​

Observations:

  1. Matrix-based optimizers including Muon & Soap consistently outperform scalar-based. Huge speedups over AdamW: 1.2-1.4x.
  2. Scalar variants (NAdamW, MARS, Cautious) are within 1.1x of AdamW after fair tuning. Small but consistent lift.
  3. Loss improvements translate to 1-2% downstream benchmark gains.

​

​

11 of 19

Phase III: Extrapolation to 1.2B

​

​

​

Fit hyperparameter scaling laws on 12 tested settings per optimizer and extrapolate to 1.2B models

​

Observations:

  1. Muon and SOAP still offer speedup over AdamW
  2. Speedup decays with model size & data size: only ~1.1x at 8x Chinchilla (~100B tokens).

​

​

12 of 19

Qualitative Observations

Early Training is Misleading

Validation-loss curves during initial LR decay tend to exaggerate performance gaps and can even reverse the eventual ranking.

Shared Norm Dynamics

Weight norm rises/falls with LR schedule. Gradient norm increases during LR decay. Consistent across optimizers. [Defazio et.al.’25]

13 of 19

Key Takeaways: Fair Optimizer Comparison

With fair tuning, many claimed 1.4-2x speedups shrink to 1.1-1.3x.

Key parameters for tuning: learning rate

Matrix-preconditioned optimizers are best-in-class, but gains decay with scale

Scalar variants provides small but consistent ~1.1x speedup

Always compare at target budget with LR decay to near zero, across scales and data ratios

14 of 19

Extension: Batch size matters empirically

Based on discussions over X. We have some more reflections on our study.

Our objective:

Find the best hyperparameters that reaches lowest loss in one pass of a fixed amount of data (this includes tuning the batch size).

In practice:

Wall time matters more, and it crucially depends on the MFUs (hardware utilization rate)

Especially, batch size positively correlates with MFUs.

Muon has much better performance compared to AdamW at large batch size.

15 of 19

Extension: Preserving the speedup

The relative slowdown of matrix optimizer at larger compute budget is very puzzling.

​

[Qiu et.al. 2025] points out that properly scaling the weight decay with 1/width can preserve the speedup of matrix-based optimizer at 1x Chinchilla regime

​

This indicates that:

Controlling the weight norm dynamics of optimizers is crucial for preserving the speedup of optimizers.

​

[Qiu et.al. 2025], Fig 1

16 of 19

Hyperball: Weight-Normalized Training

Two simple modifications:

1. Remove weight decay

2. Normalize weight and update norms

​

Compatible with any optimizer:

Adam-Hyperball (AdamH)

Muon-Hyperball (MuonH)

Why doesn't this hurt representation power?

Modern Transformers use RMSNorm with rescaling parameters. These parameters steer the output norm when weight norm is fixed.

17 of 19

Hyperball: Persistent Speedup

AdamH and MuonH show persistent 20-30% speedup over weight decay counterparts with increasing scale with AdamH even outperforms MuonW at 1.2B scale!

​

We also observe MuonH to provide constant loss gap with respect to MuonW on 130M model with extreme overtraining.

18 of 19

Hyperball: Hyperparameter transfer out-of-box

Thanks to constant ratio between update and weight norm, we observe a consistent learning rate scaling for Hyperball.

AdamH enables highly predictable scaling in Marin (0.005 difference on 8B model & 32B model on the way).

19 of 19

Applied in Current 535B Marin Open Research Training

https://wandb.ai/marin-community/marin_moe/reports/535B-A23B-18T-Token-Hero-Run-Scaling-Ladder--VmlldzoxNzc2MDM5Ng

​