AIBRIEF
Motif Technologies re released its largest open weights language model under an MIT license, letting any builder fine tune, embed, or sell products built on the weights.
Motif Technologies swapped the license on its largest open-weights language model this week, moving Motif 3 from a non-commercial research restriction to a permissive MIT license on three Hugging Face repositories. The change arrived without a press release, blog post, or notice on the company's website, according to TechTimes.
Nothing about the model itself changed. The 314-billion-parameter mixture-of-experts design (13.2 billion activated per token), the 256,000-token context window, and the proprietary GDLA attention design carry the same specs as the July beta. What changed is what a builder is allowed to do with the weights. Under the old restriction, a startup could not fine-tune Motif 3 for resale, an enterprise could not embed it in a paid product, and a hosting provider could not charge to serve it. MIT removes each of those gates.
The technical report describes pretraining on roughly 12.5 trillion tokens across web, code, mathematics, and Korean-language data, routing 8 of 384 experts plus a shared expert per token. The release lands in a Korean policy environment that bars foreign-trained weights from certain public-sector deployments. All four of the country's sovereign AI teams operate domestically trained models, and Naver Cloud and NC AI recently opted out of a sovereign-AI retender. Whether Motif 3 fills that gap depends on adoption the release does not yet claim.