Marco-MoE: Open Multilingual Mixture-of-Expert Language Models with Efficient Upcycling

TL;DR

Marco-MoE introduces a highly sparse, open multilingual mixture-of-experts model that achieves efficient training and superior performance, with scalable language expansion and shared activation patterns across languages.

Contribution

The paper presents Marco-MoE, a novel sparse multilingual MoE model that outperforms similar-sized models and enables scalable language expansion with shared expert activation patterns.

Findings

Achieves state-of-the-art performance-to-compute ratio.

Surpasses larger models in instruction tuning.

Learns shared activation patterns across related languages.

Abstract

We present Marco-MoE, a suite of fully open multilingual sparse Mixture-of-Experts (MoE) models. Marco-MoE features a highly sparse design in which only around 5\% of the total parameters are activated per input token. This extreme sparsity, combined with upcycling from dense models, enables efficient pre-training on 5T tokens. Our models surpass similarly-sized competitors on English and multilingual benchmarks, achieving a best-in-class performance-to-compute ratio. We further post-train these models to create Marco-MoE-\textsc{Instruct} variants, which surpass the performance of competing models possessing $3$--$14\times$ more activated parameters. Our analysis reveals that Marco-MoE learns structured expert activation patterns shared across related languages, while maintaining highly specialized utilization for linguistically isolated ones. We further show that Marco-MoE allows for scalable language expansion without the interference typical of dense models. To support the community, we disclose our full training datasets, recipes, and model weights.