South Korea Pushes US To Third Place In Open Models Race With Release Of Motif 3

South Korea's Motif Technologies has released Motif 3, a 314-billion-parameter open-weight AI model that scored 47 on the Artificial Analysis Intelligence Index, placing it ahead of all US open-weight models and behind only China's frontier models. The model, built by a roughly 30-person team in about five months, is part of South Korea's government-backed Dokpamo sovereign AI program and is available on Hugging Face under an MIT license.

The US isn’t just lagging China in open source AI — it’s now behind South Korea too. That’s the picture emerging from the latest Artificial Analysis Intelligence Index, where a 314-billion-parameter model called Motif 3, built by a roughly 30-person South Korean startup, has scored 47 points and slotted in ahead of every open-weight model the United States currently has on the board. China’s frontier open models, led by names like Kimi and GLM, remain well out in front at the top of the chart, above 50 and pushing toward 60. But the US’s best open offerings now sit behind South Korea’s as well, a country that most people tracking the AI race hadn’t been watching closely. Motif 3 comes from Motif Technologies, a subsidiary of Moreh, a South Korean AI infrastructure company founded in 2020 that has spent years building GPU virtualization and training software rather than models themselves. Motif Technologies was carved out as a dedicated model-development unit only in February 2025, and the team has already shipped a frontier-class open model in roughly five months of pretraining, according to the company. It raised about $16.9 million in a Series B round in May 2026 from a group of Korean investors including NICE Investment Partners and Nautilus Investment, on top of Moreh’s earlier $22 million Series B from AMD and Korea Telecom. The model itself is a Mixture-of-Experts system with 314 billion total parameters, of which only 13.2 billion are active for any given token, a design choice that keeps inference costs down while still letting the model draw on a much larger pool of learned knowledge. Motif Technologies says it built the entire pipeline, from pretraining through post-training, from scratch, and it has attributed part of the jump from the beta version’s score of 45 to two techniques it describes as proprietary: something called GDLA and a modified attention mechanism the company calls polynorm. Weights, along with training code and related materials, are up on Hugging Face under an MIT license, which is about as permissive as licensing gets and allows unrestricted commercial use. Running it is not for the casual hobbyist. Motif recommends NVIDIA B200 or H200-class hardware, and the model requires trust remote code to be enabled along with a Motif-specific vLLM Docker image to serve. For teams without access to that tier of GPU, the company has also put out a quantized NVFP4 version that trims the footprint considerably. The release is tied to something bigger than one company’s product roadmap. Motif 3 is one of four models competing in the second phase of South Korea’s government-backed sovereign AI program, known as Dokpamo, which is evaluating domestic foundation models built by SK Telecom, LG AI Research, Upstage, and Motif. On the Artificial Analysis benchmark specifically, Motif came out on top of that group by a wide margin, scoring 47 against Upstage’s Solar Open 2 at 37, SK Telecom’s A.X K2 at 35, and LG’s K-EXAONE 2.0 at 31. That benchmark carries real weight in the government’s evaluation too, since it makes up 25 of the 40 points allocated to technical benchmarking in the second-stage scoring, out of 100 points overall. What makes this notable beyond Korea’s borders is the position it puts the US in. The Artificial Analysis chart tracking open-weight frontier models by country over time shows a clear pecking order that has held for most of the past year: China running away at the top, the US as the clear second, and everyone else trailing well behind. Motif 3’s score of 47 changes that ordering, at least for now, putting a South Korean lab ahead of the best open release the US currently has out. It also fits into a broader trend that has been building for a while. Chinese labs have already established themselves as the dominant force https://officechai.com/ai/chinas-ai-models-replace-uss-ai-models-as-open-model-of-choice-for-first-time/ in open-weight models, with companies like Alibaba, DeepSeek, Moonshot, and Z.AI releasing open models that regularly top global rankings and have reportedly pulled a majority of token volume https://officechai.com/ai/share-of-us-models-being-used-on-openrouter-has-collapsed-from-70-to-30-over-the-past-year/ on platforms like OpenRouter away from US providers. OpenAI’s return to open weights with its gpt-oss models gave the US a foothold again, but the ceiling on what’s coming out of the country’s labs has stayed noticeably lower than what China is producing, a gap that Google itself hasn’t been immune to, with a Chinese open model outscoring every Gemini release https://officechai.com/ai/a-chinese-open-source-model-is-ahead-of-all-google-models-on-the-artificial-analysis-intelligence-index-for-the-first-time/ on this same index earlier this year. Meta has said that it would release the weights for its Muse 1.2 model, which scores 57 on the Artificial Analysis Intelligence Index, but hasn’t done so yet. South Korea entering the picture, and doing so credibly enough to leapfrog the US, adds a new axis to a race that had mostly been framed as a two-country contest. Motif Technologies isn’t stopping at text either. Alongside the language model, the company has been building out smaller multimodal components, a 7-billion-parameter vision module, a 0.7-billion-parameter audio module, and a 1.2-billion-parameter video VAE, which it plans to eventually combine with Motif 3 into a fuller multimodal system. For a team of around 30 people working out of a two-year-old parent company, that’s an ambitious roadmap. Whether Motif can keep pace as bigger, better-funded labs release their next generations is the real test. For now, though, the scoreboard says a small Korean startup has done something Meta, OpenAI, and the rest of the American open-source field haven’t managed this year.