Models & Weights

Motif 3

Korea's 314B sovereign MoE said MIT on the card — and now the file is in the repo to prove it

Open ✓Model

What it is

Motif Technologies' flagship: a 314B-total / 13.2B-active MoE with 384 routed experts (top-8 plus one shared), 53 layers, 256K native context, and in-house architectural signatures — Grouped Differential Latent Attention and Expert-Specific PolyNorm. The final release landed August 13 alongside Base and NVFP4 variants and a technical report (arXiv 2608.09119), with self-reported agentic numbers that punch far above the active-parameter count: 76.2 SWE-bench Verified, 83.4 GPQA Diamond, 74.9 Terminal-Bench 2.1.

Why it's interesting

Two sovereign-AI flagships from Korea in one summer, two opposite license strategies: Upstage bolted branding riders onto Apache, while Motif went plain MIT — no thresholds, no naming clause, no AUP. At 630GB of shards it's a fraction of Kimi K3's bill for benchmark claims in the same conversation. At our first check the MIT promise rested on card metadata alone — no LICENSE file in the tree, earning a nomination to our missing-LICENSE club. By the pre-publish review, the file had appeared: standard, unmodified MIT text. The asterisk this entry was drafted with is gone.

Use cases

  • Self-hosted long-horizon agents where per-token cost matters — 13.2B active is the efficiency play
  • Fine-tuning a frontier-adjacent MoE without license negotiation — MIT, file and all
  • Architecture research on differential-attention and per-expert normalization at scale

Who it's for

Infra teams wanting Kimi-class agentic scores at a third of the memory; license-sensitive builders

Setup

Advanced. 155 safetensors shards totaling ~630GB in BF16; the official NVFP4 variant is the realistic single-node path

Limitations & cautions

The headline benchmarks are self-reported and independent evals were scarce at check time, and the proprietary attention and normalization schemes mean framework support may lag standard MoEs.

Editorial takeaway

This entry was drafted ending 'if the card is telling the truth.' Between draft and publish, the file appeared — and it is. Graduation from the missing-LICENSE club costs exactly one kilobyte; the rest of the members know where the door is.

Related & alternatives