South Korea Releases Motif 3, An AI Model that Competes With Chinese Open-Source

A South Korean AI company called Motif Technologies has released Motif 3, an open-source large language model built to compete with Chinese open-source systems such as DeepSeek V4 Pro. The release is being treated as a major milestone for the country's AI ambitions, and one of the more concrete signs yet that nations outside the US and China can produce frontier class AI systems.

Motif 3 is a large scale Mixture of Experts language model built from the ground up by Motif Technologies following a fully in house, proprietary design, not a re-parameterization of existing open source architectures. That distinction matters. Rather than taking an existing open source architecture like Llama or Mistral and retraining it, Motif built its own transformer design from scratch, including some genuinely novel components. According to AI Weekly's technical breakdown, the model card describes two components not seen elsewhere at scale, a form of grouped differential latent attention and a grouped activation function applied per expert.

The model currently carries a beta label, and Motif has been upfront that this is not the finished product. This is a preview or beta checkpoint, not the final release, and the current repository hosts an intermediate checkpoint of Motif 3, with the final checkpoint expected to be released soon.

In terms of specs, the model has roughly 314 billion total parameters, with about 13 billion active per token thanks to its sparse Mixture of Experts design. Its sparse routing uses 384 experts with 8 activated per token, plus 1 shared expert. It has a 256K context length, specifically 262,144 tokens, making it natively long context. Motif 3 supports text input and outputs text, but does not support image input. It is also a reasoning model, meaning it uses extended thinking or chain of thought reasoning to work through complex problems before providing an answer.

Right now there are two main ways to use Motif 3. The first is to download the open weights and run the model yourself. The model is openly available, meaning anyone can download the weights with no access request required. However, licensing is limited, since permission is granted to use, modify, and redistribute the software for personal, educational, and non-commercial research purposes only. Commercial use requires separate written approval from Motif Technologies.

The second option, for people who do not want to host the model themselves, is Motif's own consumer facing chat product, website. It is described as a versatile AI agent for everyday moments that is smart by default, with expertise and emotional nuance built in for natural conversations. The broader platform combines the language model with other capabilities, including image generation, and is pitched as offering ChatGPT level performance for complex reasoning, summarization, translation, and code generation.

Independent evaluation firm Artificial Analysis put real numbers behind Motif's claims. Motif 3 Beta scores 44 on the Artificial Analysis Intelligence Index, placing it well above average among other reasoning models in a similar price tier, where the median score is 15. In ranking terms, that score puts it around 34th out of 577 tracked systems. One quirk worth noting is verbosity. When evaluating the Intelligence Index, the model generated 190 million tokens, which is very verbose compared to the average of 61 million. In other words, Motif 3 tends to think through problems at considerable length before landing on an answer, which likely contributes to its strong reasoning scores but also means it is slower and more compute hungry per response than many competitors.

Commentary from the AI research community has been notably positive on the technical side, with one researcher describing it as a Korean lab releasing a model with roughly 341 billion parameters that performs on par with a model nearly five times its size, and noting genuine architectural innovation behind it. AI Weekly framed the release as one of the loudest sovereign AI signals of the quarter, while also pointing out that the non-commercial license means the real question is whether Motif can convert the benchmark result into a licensable product before the moment passes.

There is currently no commercial API pricing for Motif 3, since no hosted providers have launched a paid endpoint yet. Pricing for Motif 3 Beta is listed at zero dollars per million input tokens and zero dollars per million output tokens, well below competitive averages of about 20 cents and 57 cents respectively, though this reflects the absence of a paid tier rather than a genuinely free commercial offering. Anyone wanting to run it themselves will need substantial GPU infrastructure given the model's size, and commercial deployment requires direct licensing from Motif.

Motif Technologies is a fast moving Korean startup with a clear specialty in building foundation models entirely from original architecture rather than adapting existing open source designs. The company has gained recognition for its commitment to building AI foundation models entirely from its own architecture, without borrowing from foreign open source model designs.

That approach has attracted serious government backing. Motif plans to develop a reasoning focused large language model with 300 billion parameters built on domestically developed AI technology, and with its selection, Motif joins LG AI Research, Upstage and SK Telecom in South Korea's sovereign AI foundation model project. The Ministry of Science and ICT specifically praised the company's originality, noting Motif was chosen for its experience designing models with its own architecture and for achieving performance it said could compete with leading global systems despite operating in a limited data environment.

Motif has also secured fresh capital to keep building. The company has closed an approximately 16.6 million dollar, or 24 billion Korean won, Series B funding round. Its roadmap goes beyond text only models, with plans calling for the LLM to evolve progressively into a vision language model with around 310 billion parameters and ultimately a vision language action model with around 320 billion parameters, positioning Motif Technologies to compete at the global frontier of AI foundation model development.

Motif 3 is significant less because of raw benchmark numbers and more because of what it represents, proof that a relatively small, independent company outside the US and China can design a competitive frontier scale model from first principles. Given the current landscape of the AI race being dominated by American labs like OpenAI, Google, and Anthropic, and Chinese labs like DeepSeek, Alibaba's Qwen, and Moonshot's Kimi, a Korean startup landing a top forty global ranking with a homegrown architecture is a meaningful data point for the broader sovereign AI movement, where countries are increasingly investing in independent AI capabilities rather than relying solely on foreign models. The open question now is whether Motif can turn this technical achievement into a sustainable commercial business once the final, non beta version of Motif 3 ships.