Motif-3-Beta
Released Jul 20, 2026 · knowledge cutoff unpublished
- Status
- Superseded
- Location
- South Korea
- Modality
- Text
- Context window
- 262K
- Max output
- —
- Speed
- —
- Price ($/MTok in / out)
- — / —
- Cost per task
- —
- Open weights
- Yes
Benchmarks3 of 8 reported
GPQA Diamond
Terminal-Bench 2.1
Humanity's Last Exam
MMLU-Pro
SWE-bench Verified
AIME (latest)
LMArena Elo
ARC-AGI-2
Strengths & weaknesses
Strengths
- A from-scratch architecture rather than a re-parameterised open model — its grouped differential latent attention learns a per-group noise head and subtracts it back out
- Carries a multi-token-prediction head, so it self-speculates without a separate draft model
- Korean is a first-class training target, not a translation afterthought
- Weights are ungated — no access request, no waitlist
Weaknesses
- An explicitly intermediate checkpoint with the final Motif-3 still to come, so its behavior is a moving target
- Non-commercial licence only, a step back from Motif-2's Apache terms, which rules it out of production
- Custom modelling code needs trust_remote_code and a vendor vLLM image tested only on B200 and H200 — anything else is on you
Preview checkpoint from Motif Technologies, a ~30-person Moreh subsidiary building under South Korea's sovereign-AI programme; weights landed on Hugging Face 2026-07-20 and Artificial Analysis dates the checkpoint 2026-07-14. Motif publishes no benchmark table of its own beyond an Artificial Analysis Intelligence Index of 44, so GPQA Diamond 86.9, Terminal-Bench 2.1 70.8 and HLE 38.2 here are Artificial Analysis' own measurements; MMLU-Pro, SWE-bench, AIME, LMArena and ARC-AGI-2 are unpublished. No per-token price exists and no aggregator serves it, so cost per task is unset — the only hosted route is a free chat at chat.motiftech.io, hence availability 'restricted' despite downloadable weights.
For developers
API model strings
Not researched
Licence
Not researched
Retirement
No retirement announced
Lineage
No recorded predecessor