HUMAIN unveils a 428B Arabic model

HUMAIN’s humain-m3 is a 428-billion-parameter mixture-of-experts model trained on more than one trillion Arabic-native tokens and now available in research preview across seven public Arabic benchmarks. HUMAIN’s announcement says weights are targeted for release next month under the MiniMax Community License, not yet generally available. Teams serving Arabic users should benchmark local-language retrieval and safety before assuming an English-first frontier model is sufficient, alongside the archive’s case for small open weights and control.