We analyze Muon as originally proposed and used in practice -- using the momentum orthogonalization with a few Newton-Schulz steps. The prior theoretical results replace this key step in Muon with an exact SVD-based polar factor. We prove that Muon with Newton-Schulz converges to a stationary point at the same rate as the SVD-polar idealization, up to a constant factor for a given number $q$ of Newton-Schulz steps. We further analyze this constant factor and prove that it converges to 1 doubly exponentially in $q$ and improves with the degree of the polynomial used in Newton-Schulz for approximating the orthogonalization direction. We also prove that Muon removes the typical square-root-of-rank loss compared to its vector-based counterpart, SGD with momentum. Our results explain why Muon with a few low-degree Newton-Schulz steps matches exact-polar (SVD) behavior at a much faster wall-clock time and explain how much momentum matrix orthogonalization via Newton-Schulz benefits over the vector-based optimizer. Overall, our theory justifies the practical Newton-Schulz design of Muon, narrowing its practice-theory gap.

The Muon optimizer, a matrix-structured algorithm that leverages spectral orthogonalization of gradients, is a milestone in the pretraining of large language models. However, the underlying mechanisms of Muon -- particularly the role of gradient orthogonalization -- remain poorly understood, with very few works providing end-to-end analyses that rigorously explain its advantages in concrete applications. We take a step by studying the effectiveness of a simplified variant of Muon through two case studies: matrix factorization, and in-context learning of linear transformers. For both problems, we prove that simplified Muon converges linearly with iteration complexities independent of the relevant condition number, provably outperforming gradient descent and Adam. Our analysis reveals that the Muon dynamics decouple into a collection of independent scalar sequences in the spectral domain, each exhibiting similar convergence behavior. Our theory formalizes the preconditioning effect induced by spectral orthogonalization, offering insight into Muon's effectiveness in these matrix optimization problems and potentially beyond.

The growing scale of deep learning models has rendered standard hyperparameter (HP) optimization prohibitively expensive. A promising solution is the use of scale-aware hyperparameters, which can enable direct transfer of optimal HPs from small-scale grid searches to large models with minimal performance loss. To understand the principles governing such transfer strategy, we develop a general conceptual framework for reasoning about HP transfer across scale, characterizing transfer as fast when the suboptimality it induces vanishes asymptotically faster than the finite-scale performance gap. We show formally that fast transfer is equivalent to useful transfer for compute-optimal grid search, meaning that transfer is asymptotically more compute-efficient than direct tuning. While empirical work has found that the Maximal Update Parameterization ($μ$P) exhibits fast transfer when scaling model width, the mechanisms remain poorly understood. We show that this property depends critically on problem structure by presenting synthetic settings where transfer either offers provable computational advantage or fails to outperform direct tuning even under $μ$P. To explain the fast transfer observed in practice, we conjecture that decomposing the optimization trajectory reveals two contributions to loss reduction: (1) a width-stable component that determines the optimal HPs, and (2) a width-sensitive component that improves with width but weakly perturbs the HP optimum. We present empirical evidence for this hypothesis across various settings, including large language model pretraining.

Training large language models (LLMs) typically relies on adaptive optimizers such as Adam, which introduce extra operations and require significant more memory to maintain first- and second-order moments than SGD. While recent works such as GaLore, Fira and APOLLO have proposed state-compressed variants to reduce memory consumption, a fundamental question remains: What are the minimum modifications to plain SGD needed to match state-of-the-art pretraining performance? We systematically investigate this question using a bottom-up approach, and identify two simple yet highly (memory- and compute-) efficient techniques: (1) column-wise gradient normalization (normalizing the gradient along the output dimension), which boosts SGD performance without momentum; and (2) applying first-order momentum only to the output layer, where gradient variance is highest. Combining these two techniques lead to SCALE (Stochastic Column-normAlized Last-layer momEntum), a simple optimizer for memory efficient pretraining. Across multiple LLaMA models (60M-1B), SCALE matches or exceeds the performance of Adam while using only 35-45% of the total memory. It also consistently outperforms memory-efficient optimizers such as GaLore, Fira and APOLLO, making it a strong candidate for large-scale pretraining under memory constraints. For LLaMA 7B model, SCALE outperforms the state-of-the-art memory-efficient methods APOLLO and Muon, in terms of both perplexity and memory consumption.

In this article, we explore the use of various matrix norms for optimizing functions of weight matrices, a crucial problem in training large language models. Moving beyond the spectral norm underlying the Muon update, we leverage duals of the Ky Fan $k$-norms to introduce a family of Muon-like algorithms we name Fanions, which are closely related to Dion. By working with duals of convex combinations of the Ky Fan $k$-norms with either the Frobenius norm or the $l_\infty$ norm, we construct the families of F-Fanions and S-Fanions, respectively. Their most prominent members are F-Muon and S-Muon. We complement our theoretical analysis with an extensive empirical study of these algorithms across a wide range of tasks and settings, demonstrating that F-Muon and S-Muon consistently match Muon's performance, while outperforming vanilla Muon on a synthetic linear least squares problem.

We report on the first large-scale mixture-of-experts (MoE) pretraining study on pure AMD hardware, utilizing both MI300X GPUs and Pollara networking. We distill practical guidance for both systems and model design. On the systems side, we deliver a comprehensive cluster and networking characterization: microbenchmarks for all core collectives (all-reduce, reduce-scatter, all-gather, broadcast) across message sizes and GPU counts over Pollara. To our knowledge, this is the first at this scale. We further provide MI300X microbenchmarks on kernel sizing and memory bandwidth to inform model design. On the modeling side, we introduce and apply MI300X-aware transformer sizing rules for attention and MLP blocks and justify MoE widths that jointly optimize training throughput and inference latency. We describe our training stack in depth, including often-ignored utilities such as fault-tolerance and checkpoint-reshaping, as well as detailed information on our training recipe. We also provide a preview of our model architecture and base model - ZAYA1 (760M active, 8.3B total parameters MoE, available at https://huggingface.co/Zyphra/ZAYA1-base) - which will be further improved upon in forthcoming papers. ZAYA1-base achieves performance comparable to leading base models such as Qwen3-4B and Gemma3-12B at its scale and larger, and outperforms models including Llama-3-8B and OLMoE across reasoning, mathematics, and coding benchmarks. Together, these results demonstrate that the AMD hardware, network, and software stack are mature and optimized enough for competitive large-scale pretraining.

Large language models deliver strong reasoning and tool-use skills, yet their computational demands make them impractical for edge or cost-sensitive deployments. We present \textbf{Xmodel-2.5}, a 1.3-billion-parameter small language model designed as a \emph{drop-in agent core}. Training with maximal-update parameterization ($μ$P) allows hyper-parameters tuned on a 20M-parameter proxy to transfer directly to the full model, even under the parameter-tied \emph{tie-word-embedding} architecture. A 1.4T-token Warmup--Stable--Decay curriculum is used, and we further show that \textbf{switching from AdamW to Muon during the decay phase} improves the 13-task reasoning average by 4.58\,\% while keeping every other hyper-parameter fixed, verifying that early AdamW stability can be paired with late Muon sharpening for better downstream performance. FP8-mixed-precision training balances accuracy and throughput. All checkpoints, recipes, and evaluation code are released under the Apache-2.0 license.\footnote{https://huggingface.co/XiaoduoAILab/Xmodel-2.5 and https://huggingface.co/XiaoduoAILab/Xmodel-2.5-history (training checkpoints).} Training code and evaluation harness: https://github.com/XiaoduoAILab/Xmodel-2.5.

Muon, a recently proposed optimizer that leverages the inherent matrix structure of neural network parameters, has demonstrated strong empirical performance, indicating its potential as a successor to standard optimizers such as AdamW. This paper presents theoretical analysis to support its practical success. We provide convergence proofs for Muon across four practical settings, systematically examining its behavior with and without the inclusion of Nesterov momentum and weight decay. Our analysis covers the standard configuration using both, thereby elucidating its real-world performance. We then demonstrate that the addition of weight decay yields strictly tighter theoretical bounds and clarify the interplay between the weight decay coefficient and the learning rate. Finally, we derive the critical batch size for Muon that minimizes the computational cost of training. Our analysis identifies the hyperparameters governing this value, and our experiments validate the corresponding theoretical findings across workloads including image classification and language modeling task.

Abstract--Muon optimizer has demonstrated robust results in pretraining of language models but its performance in finetuning of existing public pretrained models is not yet explored. Currently, Muon is used along with AdamW introducing a scope of improvement for adopting all parameters inside Muon. We introduce MuonAll, which incorporates all the parameters inside Muon by transforming into 2D matrices. We conduct extensive finetuning experiments across publicly available language models with model sizes upto half billion parameters. Muon and MuonAll perform at par with AdamW across major benchmarks, highlighting their effectiveness as alternative optimizers.

In this paper, we introduce a model for analyzing deep learning optimization over a single iteration by leveraging the matrix structure of the weights. We derive the model by assuming isotropy of curvature, including the second-order Hessian and higher-order terms, of the loss function across all perturbation directions; hence, we call it the isotropic curvature model. This model is a convex optimization program amenable to analysis, which allows us to understand how an update on the weights in the form of a matrix relates to the change in the total loss function. As an application, we use the isotropic curvature model to analyze the recently introduced Muon optimizer and other matrix-gradient methods for training language models. First, we show that under a general growth condition on the curvature, the optimal update matrix is obtained by making the spectrum of the original gradient matrix more homogeneous -- that is, making its singular values closer in ratio -- which in particular improves the conditioning of the update matrix. Next, we show that the orthogonalized gradient becomes optimal for the isotropic curvature model when the curvature exhibits a phase transition in growth. Taken together, these results suggest that the gradient orthogonalization employed in Muon and other related methods is directionally correct but may not be strictly optimal. Finally, we discuss future research on how to leverage the isotropic curvature model for designing new optimization methods for training deep learning and language models.