With the release of China's Moonshot AI's Kimi K3 and the 'Open Weight and U.S. AI Leadership' letter proposed by NVIDIA CEO Jensen Huang, the global AI market is rapidly moving from a focus on closed models to competition on open models. In Korea, domestic open weight models such as Solar Open 2 and Motif 3 emerged and responded through its own AI foundation business. In the future, it is important to not only select a national representative model, but also connect all domestic AI companies to secure technological sovereignty and ecosystem competitiveness.

1. Global AI shock triggered by Kimi K3

On July 16, 2026, Chinese AI startup Moonshot AI unveiled its super-large AI model 'Kimi K3'. Kimi K3 is a model of mixture-of-experts, or MoE Structure, with 2.8 trillion parameters.

Unlike existing companies such as OpenAI and Anthropic, which do not disclose core models and provide them as paid services on their own servers, Kimi K3 has completely disclosed model weights. Accordingly, instead of accessing Moonshot AI's service, users can download the model directly and run it on their own server. This is significant in that it increases the way the model works and its usability, and reduces the problem of users being dependent on services from specific companies.

Performance was also evaluated at a significant level. Kimi K3 showed performance equivalent to top 4 in the world following the latest models such as Claude Opus 5, Fable 5, and GPT-5.6 Sol. Although it has the disadvantages of being somewhat verbose, slow in speed, and high in token creation costs, it shocked the industry in that Chinese AI quickly approached the level of American AI.

"The intelligence of Chinese AI quickly caught up with American AI."

While DeepSeek showed last year that high-performance AI models can be developed cheaply and efficiently, Moonshot AI proved that large startups can also create super-large AI with high intelligence. This has increased the sense of crisis that the AI ​​technology gap between the United States and China is rapidly narrowing.

In this atmosphere, on July 24, NVIDIA CEO Jensen Huang proposed the 'Open Weight and U.S. AI Leadership' Letter. Major technology companies including NVIDIA, Intel, AMD, Amazon, Microsoft, Meta, and IBM participated in this letter, and within a week, 130 companies signed.

2. Why the open weight model is attracting attention

Open Weight refers to an open model that discloses the weight, which is a key parameter of the AI model. The opposite concept, Closed Model, is a method in which a company provides services only through a web page or API without disclosing the model's parameters and learning structure.

The closed model is convenient because users do not need to prepare separate infrastructure and companies continuously update the model. However, there are limitations in that it is difficult to secure data sovereignty because it requires dependence on the service provider and it is difficult to control where and how the user's data is processed.

On the other hand, open weight models allow users to download model weights and run them on their own servers or devices. Companies or research institutes can operate models on their own without continuing to pay for external AI services, and AI can be utilized even in environments where it is difficult to use external services, such as financial institutions or closed networks.

However, to use the open-weight model, you must build and manage the server and GPU infrastructure that will run the model. Additionally, there is the possibility of disabling or modifying the model's safety features, complicating liability issues for abuse and accidents.

Surprisingly, Google and OpenAI were also named in the current Open Weight and US AI leadership letters. Both companies are expanding the influence of the AI ​​ecosystem in various ways by releasing not only closed models but also open source and open weight models. Anthropic, on the other hand, did not sign the letter.

Anthropic CEO Dario Amodei stated that he does not oppose open-weight per se, but explained that he does not agree with the argument that unrestricted AI accessibility could actually provide greater benefits to attackers.

"I am not against the open weight model itself, and the non-dangerous open weight model is a public good."

"I agree with much of Open-weight and the U.S. AI leadership letter, but I disagree with the argument that broad AI accessibility benefits defenders over attackers."

What Anthropic is concerned about is not just the business burden. This is because if the open weight model is released, copyright issues with learning data, controversies over who should be responsible between the model manufacturer and operator in the event of an accident, and the possibility of abuse bypassing safety devices may increase.

3. Independent AI foundation business that became a domestic response measure

If Open Weight becomes a new standard in the global AI market, major changes are expected in the domestic industry. Companies and research institutes that have so far used overseas closed models such as GPT or Claude can directly introduce overseas open weight models by providing their own infrastructure. A big advantage is that high-performance AI can be operated without sending data to the outside world, even in financial or closed network environments.

However, if the domestic model does not secure sufficient competitiveness, there is a possibility that the domestic market will be quickly encroached upon by the open weight model of overseas big tech. At this time, an important countermeasure is the Independent AI Foundation Project, which started last year.

This project is a policy to secure our own AI-based model made with domestic infrastructure and technology, rather than relying on overseas big tech models. The goal is to have performance of 95% or more compared to global leading AI models.

In addition, a key condition is to disclose the model's weights, architecture, and learning methodology so that anyone can use it in various fields such as domestic industry, education, and science. Through this, the goal is to create an ecosystem where domestic companies and research institutes can freely utilize and improve models, rather than monopolizing AI technology for certain companies.

4. Technical achievements of Solar Open 2 and Motif 3

As of July 28, 2026, the open AI models receiving the most attention in Korea are Upstage's Solar Open 2 and Motif Technologies' Motif 3. LG AI Research Institute's K-ExaOne 2.0 and SK Telecom's A.X K2 are also scheduled to be released soon.

Models participating in the independent AI foundation project must disclose not only the weights but also the architecture and learning methodology through hugging faces, etc., so it is expected that various open source and open weight models will appear in Korea in the future.

Solar Open 2: Focus on handling agents and extremely long sentences

On July 22, Upstage unveiled the open weight model Solar Open 2 on Hugging Face. While the previous generation Solar Open 100B focused on general-purpose large language model performance, Solar Open 2's main goal was actual agent AI utilization such as tool calls, coding, and office work.

The model scale is 250 billion parameters. Users can download and use weights, and can also create commercial services or derived AI models by including 'Solar' in the model name and indicating 'Built with solar' in the source.

Performance has also improved significantly compared to the previous version.

• MMLU-Pro recorded higher performance than overseas models such as DeepSeek V4-Flash and Mimo-V2.5.
• LiveCodeBench v6 accuracy has increased from 56.5% of the previous version to 92.4%.
• APEX-Agent, which evaluates the work of financial and management experts, also significantly outperformed Mistral Medium 3.5 and DeepSeek V4 Flash.
• It also received the highest scores in KMMLU-Pro, K-Knowledge, K-Domain, and K-Office Agent, which are specialized for Korean language and domestic work.

Technically, it combines Linear Attention, which efficiently summarizes long texts, and Softmax Attention, which is responsible for detailed context understanding, at a ratio of 75 to 25. Through this, it was designed to handle long contexts while lowering the memory burden.

Additionally, out of a total of 250 billion parameters, only 15 billion are activated during the actual answer process. Out of a total of 320 optional experts, 8 and 1 common expert operate according to the question. Thanks to this, the overall size of the model is large, but the amount of computation in the actual inference process can be reduced.

Solar Open 2 significantly reduced the number of tokens that Solar Open 100B needed to reach a certain performance. For example, while Solar Open 100B required 671 billion tokens to achieve MMLU 0.55, Solar Open 2 reached the same level with only 210 billion tokens.

Driving conditions have also been improved. It can be run with 4 NVIDIA H200 GPUs based on 16-bit floating point, and if quantization is applied, it can also be run with 2 H200 GPUs. Nota applied NVFP4 quantization for NVIDIA Blackwell to reduce the model size from 500.6GB to 153.3GB.

However, some say that it is not a model with a fast inference speed because the multi-token prediction technology MTP is not applied. However, Solar Open 2 is a model that focuses on extremely long context processing of more than 1 million tokens and agent execution ability rather than speed, and is interpreted as prioritizing core functions within limited computing resources.

Motif 3: Proprietary architecture and high efficiency

Motif Technologies released Motif 3 to Hugging Face on July 21 as part of its independent AI foundation business. Motif began development through an additional public offering in February of last year, and completed a frontier-class model in 5 months** with about 30 people and 700 GPUs.

Motif 3 has 314 billion parameters, but during actual inference, only 13 billion, or about 4% of the total, are activated. When a question comes in, 8 experts are selected out of 384 experts and 1 common expert to create an answer.

GLDA, or grouped differential latent attention structure, was applied to this model. This is a method that combines differential attention to reduce irrelevant words, latent attention to save memory by compressing the KV cache into a vector, and grouping technology to manage the parallel processing and resource allocation of the two technologies.

Another characteristic of Motif 3 is that the entire development process, starting with the architecture, was designed with proprietary technology rather than combining existing open sources.** Using open source can shorten the development period, but it may be affected by changes in the original author's policy or restrictions on commercialization. Conversely, proprietary structures allow you to directly optimize your model to fit the needs of a specific industry or company.

Performance was also highly evaluated. Motif 3 scored 44 points in Artificial Analysis' IQ. This figure significantly exceeds the average median score of 15 points for similar models, and ranks 40th among all 586 models, including duplicate scores, and 3rd among open source models.

The current IQ was evaluated at the level of DeepSeek V4 Pro and Claude Opus 4.6. Considering that it is still a pre-release beta model, there is a possibility that performance will be further improved in the final version.

5. Direction of the domestic AI ecosystem

With the release of Kimi K3 and the Open Weight letter, the global AI market is likely to be reorganized into a competition between ultra-high-performance closed models and open weight models that emphasize price-performance ratio.

If big American tech companies actively introduce the open weight model, companies, governments, and research institutes around the world will utilize it. In the short term, it can lower usage costs and increase technology accessibility, but in the long term, there is a risk of being dependent on the AI ​​ecosystem of a specific country or company.

In this situation, to protect technological sovereignty, we must secure a high-performance foundation model developed domestically. The independent AI foundation business goes beyond simply creating domestically produced AI that the public can use, and can become the basis for reducing dependence on foreign models and maintaining the domestic ecosystem.

However, by the end of this year, only two of the four models currently competing will be selected as national representative AI. Although it can be efficient in terms of selection and focus, it can be risky in that it requires securing as many technological options as possible in the global AI competition.

"The independent AI foundation business must be of a nature that goes beyond selecting and focusing on Korean-style AI and connects the entire domestic AI ecosystem."

Therefore, rather than leaving only two models in the end, there is a need to connect all domestic AI companies and consortiums that have participated in the project so far and develop them in a way that shares technology and infrastructure. This is because not only the representative model, but also the models and technologies of various companies must grow together to maintain independent technological sovereignty in the future global AI hegemony competition.

Conclusion

The spread of the open weight model is changing the era in which high-performance AI was monopolized by only a few big tech companies, and paving the way for companies and governments to operate AI in their own infrastructure. However, as openness increases, issues such as safety, copyright, and liability must also be addressed.

In Korea, Solar Open 2 and Motif 3 are already showing meaningful performance and efficiency. In the future, if the independent AI foundation business expands to an ecosystem strategy that connects all domestic companies rather than just selecting a specific model, Korea is expected to be able to establish a breakwater that protects technological sovereignty in the global open weight competition.