~314B total / ~13B active per token (sparse MoE)
262,144 tokens
Custom non-commercial research license
Official preview checkpoint of Motif-3 from South Korea's Motif Technologies, positioned as a fully in-house large-scale sparse MoE model rather than a repackaged derivative of an existing open architecture. The public Hugging Face model card describes Motif-3-Beta as an intermediate checkpoint ahead of the final Motif-3 release, with about 314B total parameters, roughly 13B active parameters per token, 384 routed experts with top-8 activation plus 1 shared expert, 53 layers, and a native 256K context window. Motif says the architecture introduces several custom components, including Grouped Differential Latent Attention (GDLA), Grouped PolyNorm activations applied per expert, and a modified mHC module. The release is framed as a multilingual general-purpose long-context model and the model card highlights an Artificial Analysis Intelligence Index score of 44, placing it in a serious open-model capability band despite being explicitly labeled as a beta rather than the final checkpoint. Public usage guidance covers Transformers, vLLM, SGLang, and Docker-based serving, while the repo notes that a dedicated vLLM serving guide is still forthcoming and community reports already indicate that the shipped custom modeling code remains rough around the edges. The current release therefore reads as an unusually large public beta artifact from Motif's in-house Korean sovereign-model effort: openly downloadable, architecturally distinctive, but still unfinished enough that the team warns the final checkpoint will follow soon.