Motif 3
#46 in Open-Source LLMsMotif Technologies · v3 · 2× · last seen Sep 04, 2026
Motif 3 is a completely in-house developed, Open Source Large Language Model with Mixture-of-Experts architecture created by Motif Technologies (a Korean startup, spin-off from Moreh). It features 314 billion total parameters, with 13.2 billion active per Token. The model employs a proprietary architecture with Grouped Differential Latent Attention (GDLA) and was trained on approximately 12.5 trillion Tokens, with special focus on Korean, Reasoning, law, and finance. In the Artificial Analysis Intelligence Index Benchmark, Motif 3 achieved 47 points and is considered the leading model among Korean language models within the government "Dokpamo" initiative, significantly ahead of competitors such as Upstage, SK Telecom, and LG AI Research.
Features
| Key Benchmark (%) | 47 points on the Artificial Analysis Intelligence Index (AAII); highest score among Korean Dokpamo models, ahead of Upstage (37) |
| Context Window (Tokens) | 262,144 tokens (256K) |
| License | MIT License |
| Multimodality | Text only (input and output), no image/audio/video support |
| Platform | Hugging Face (weights download), vLLM fork for serving, API early access via motiftech.io |
| Price per 1M Tokens | $0.00 input / $0.00 output (open-weight model, self-hostable; no official API pricing published) |
| Release Date | August 12, 2026 (final Motif 3 release) |