Open Models Almanac
Methodsoptimizationoptimizer

specific method · filed under optimization

Muon

Muon is the matrix-based optimizer used for matrix parameters, with several groups describing its general use rather than a distinct variant.

The sources on this page name it without explaining how it works. The line above is the extraction's one-line gloss. See optimizer for the methods filed beside it.

Also called Muon optimizer, Muon matrix-based optimizer.

sources
13
models
8
labs adopt it
7
strongest
core

Documented in

Alibaba (Qwen)
Poolside
Aleph Alpha
Mistral AI
Moonshot AI
Z.ai (Zhipu AI)

Evidence

15 quotes from 13 sources, one card per model, each opening with its strongest.

How sources treat it
used 12core 3

DeepSeek-V4-Pro

DeepSeekadopted4 sources

We employ the Muon optimizer for the majority of modules in DeepSeek-V4 series

coreoptimizationin DeepSeek-V4 seriesDeepSeek
3 more from this model

Muon Optimizer: We employ the Muon optimizer for faster convergence and greater training stability.

usedoptimizationin DeepSeek-V4 seriesDeepSeek

Muon Optimizer: employs the Muon optimizer for faster convergence and improved training stability.

usedoptimizationin DeepSeek-V4DeepSeek

the Muon optimizer

usedoptimizationin DeepSeek-V4DeepSeek

DeepSeek-V4-Flash

DeepSeekadopted3 sources

We employ the Muon optimizer for the majority of modules in DeepSeek-V4 series

coreoptimizationin DeepSeek-V4 seriesDeepSeek
2 more from this model

Muon Optimizer: employs the Muon optimizer for faster convergence and improved training stability.

usedoptimizationin DeepSeek-V4DeepSeek

employ the Muon optimizer for faster convergence

usedoptimizationin DeepSeek-V4-FlashDeepSeek

Qwen3.8-Flash-Next

Alibaba (Qwen)adopted2 sources

We use Muon as the main optimizer

coreoptimizationin Qwen3.8-NextQwen Team
2 more from this model

Muon is applied to the two-dimensional weights that act as linear maps

usedunclearin Qwen3.8-Flash-NextQwen Team

The Muon optimizer is used

usedoptimizationin Qwen3.8-Flash-NextQwen Team

Mistral-Large-4

Mistral AIadopted1 source

We use Muon [38, 39] (Nesterov momentum 0.95, quintic Newton–Schulz with 10 iterations, spectral scaling) at a constant learning rate of 8×10 −7

usedunclearin Mistral Large 4Mistral AI

Kimi K3

Moonshot AIadopted1 source

Following Kimi K2, Kimi K3 adopts Muon [53] as the optimizer for its matrix parameters.

usedoptimizationin Kimi K3Moonshot AI

Further reading

Picked by hand, not extracted: where to read more, not evidence for anything on this page.

↑