Motif Technologies

Motif-3-Beta

구독하기

구독하기

Motif-3-Beta is a preview (beta) checkpoint of a large-scale MoE language model designed fully in-house by Korean AI startup Motif Technologies. Rather than borrowing an existing open-source architecture, it was built from scratch with proprietary components such as Grouped Differential Latent Attention (GDLA), Grouped PolyNorm activation applied per expert, and a Multi-Token Prediction head that enables self-speculative decoding. About 13B of 314B total parameters activate per token, with 53 layers routing 384 experts (top-8) plus one shared expert, and a 262,144-token context is supported natively. It scored 44 on the Artificial Analysis Intelligence Index, placing it in the upper tier of tracked models worldwide, and is served via a custom vLLM Docker image validated on B200 and H200 GPUs. The weights are openly downloadable without access requests, but the license is limited to personal, educational, and non-commercial research (commercial use requires written permission), with a final checkpoint to follow.

전문가 성능 평가

관련 정보

PrivateAI Native

Motif Technologies

🇰🇷 Seoul2025~30