Motif-Technologies has officially released its Motif-3 large language model as part of South Korea's AI Foundation Model project (독파모). The 314B-parameter mixture-of-experts model is positioned against competitors Upstage, LG AI Research, and SKT in the initiative's second round.

  • Architecture: 314B total parameters with 13B active parameters per token (314B-A13B).
  • Agentic benchmarks: Scores include 38.7 on GDPVal v2, 94.7 on τ²-Bench Telecom, and 51.5 on ITBench.
  • Coding benchmarks: Achieved 76.2 on SWE-Bench Verified and 74.9 on Terminal-Bench 2.1.
  • Reasoning and knowledge: Recorded 83.4 on GPQA Diamond, 83.2 on IMOAnswerBench, and 75.5 on Apex-Shortlist.
  • Long context: Obtained 78.2 on IFBench and 72.3 on AA-LCR.

The release allows Motif to compete directly with other major Korean foundation models in the national project's evaluation rounds.