Motif-3
Released Aug 13, 2026 · knowledge cutoff unpublished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- Superseded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 10 reported
GPQA Diamond
SWE-bench Verified
Terminal-Bench 2.1
MMLU-Proretired
SWE-bench Pro
AIME (latest)retired
Humanity's Last Exam
LMArena Elo
GDPval-AA v2
ARC-AGI-2
Strengths & weaknesses
Strengths
- MIT-licensed where the beta was research-only — the first Motif-3 checkpoint that can legally go into a product
- Ships with a benchmark table of its own this time, and it leans agentic — tool-use and telecom-style multi-step tasks are where Motif says it does best
- Declines rather than invents when it doesn't know — its card leads with one of the higher non-hallucination rates on AA-Omniscience
- Built from scratch under Korea's sovereign-AI rules with Korean as a first-class training target, and released as base, instruct and NVFP4 checkpoints
Weaknesses
- Every score is Motif's own — no third party has reproduced the card, and Artificial Analysis' index moved only two points over the beta
- Nobody serves it — no inference provider or aggregator lists it, so the only route is self-hosting on B200/H200-class hardware with trust_remote_code and Motif's own vLLM image
- Weights appeared on Hugging Face days before any announcement and the site carried no launch post — documentation beyond the card and technical report is thin
The final Motif-3 release from Motif Technologies (a ~30-person Moreh subsidiary in South Korea's Dokpamo sovereign-AI programme): same 314B-total / 13.2B-active MoE (384 routed experts, 8 active) as the July beta, now MIT-licensed, published as Motif-3-Base, Motif-3 (instruct) and Motif-3-NVFP4 with a technical report (arXiv 2608.09119). Weights appeared on Hugging Face around 2026-08-10 unannounced; Motif's announcement followed on 2026-08-13, the date used here. SWE-bench Verified 76.2, Terminal-Bench 2.1 74.9 and GPQA Diamond 83.4 are Motif's own model-card figures (temperature 1.0, top-p 0.95, 262,144-token max sequence length), alongside τ²-Bench Telecom 94.7 and AA-Omniscience non-hallucination 71.6; none are independently reproduced. Artificial Analysis scores the final release 47 on its Intelligence Index (beta: 45) but its per-benchmark breakdown, speed and cost figures could not be read this session. HLE, MMLU-Pro, AIME, LMArena and ARC-AGI-2 are not in the card. No inference provider hosts it and Motif publishes no per-token price, so pricing and cost per task are unset and availability is 'restricted' (free chat at chat.motiftech.io or self-host) despite downloadable weights. Motif's card does not state that this release replaces the beta, so predecessorId stays null.
For developers
API model strings
Not researched
Licence
MIT License · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor