Within a few days, the reaction reached the financial markets as well: shares of technology and semiconductor companies worldwide fell, and investors began talking about the second “DeepSeek” moment – an analogy to early 2025, when a model created by another Chinese company wiped nearly $600 billion off Nvidia’s market value in one day.
Read more France and Germany plan to jointly develop an action plan on trade with China
However, the most important message of this story is not found in stock charts. It is much simpler and at the same time more significant: top-level AI is no longer the property of a few closed US laboratories. This changes the rules of the game for everyone, including Lithuania.
What really happened
“Kimi K3” is not just another model in a long string of news. In the independent “Artificial Analysis” AI model ranking, it ranks fourth among 189 evaluated models – the highest position ever achieved by an open-weight model, second only to the most powerful closed flagships. In some areas, such as long-term information retrieval and solving complex programming tasks, it outperforms all competitors, and in blind human evaluation tests by Arena.ai, designed for internet interface creation, it took first place.
Price is no less important. Using “Kimi K3” via the programming interface costs about three times less than the nearest closed competitor. Moreover, from July 27, the company promises to release all model weights. This means that any organization will be able to download, modify, and use the model on their own infrastructure without paying a cent in licenses.
Still, it is worth maintaining a sober perspective: some results are currently published by the manufacturer itself, and independent evaluations are still catching up. However, the overall direction is clear even without them – the gap between open and closed models, which seemed insurmountable just two years ago, has shrunk to a few percentage points.
Open weights are not yet open source
Here is an important distinction often lost in public discussions – similar to the difference between AI usage and AI impact. “Kimi K3” is an open-weight model but not open source by the strict standards of that term. The model weights will be downloadable and usable, but the training data, source code, and much of the technical solutions will remain the company’s secret. In other words, we get a powerful engine but not its blueprints.
This is a deliberate business model also used by other Chinese laboratories: model weights are released to the market to spread the technology as widely as possible, while the methods ensuring competitive advantage are kept within the company. For users, this means a dual situation – the freedom to use and adapt the model to their needs, but limited ability to verify what is inside and how it was created.
Openness has become a geopolitical tool
It is no coincidence that “Kimi K3” appeared on the eve of the global AI conference held in Shanghai. There, Chinese leadership presented open-source AI as a universal good and an alternative to closed Western models. The US reaction was stormy: some accuse Chinese companies of “distilling” American models, others call for regulation to limit the use of Chinese open models, and some warn that such bans would be the real defeat in the AI race.
Europe found itself in the middle of this discussion and probably the wave of open models is most relevant to it. No European company competes in model development with US or Chinese giants, and dependence on closed foreign platforms is increasingly identified as a strategic risk. Therefore, the European Commission included open source in its technological sovereignty strategy this year, and major European companies – from Siemens to Renault – are already consciously combining American, Chinese, and European models to avoid dependence on a single provider whose services can be shut down at any time.
Read more “airBaltic” airlines – bad news from Estonia: consumer rights included in the blacklist
What it means for Lithuania
This wave gives Lithuania an opportunity that did not exist a few years ago. When a top-level model can be downloaded and managed on your own infrastructure, the small size of the country ceases to be a decisive disadvantage. There is no need for billion-dollar investments to create a model from scratch – what is needed is the competence to evaluate, adapt, and safely implement it. And that is a matter of people’s competence and discipline, not scale.
The practical directions are quite clear. Open models allow creating solutions where sensitive data never leaves the organization or country – this is especially relevant for the public sector, healthcare system, and financial institutions. They also allow adapting the technology to the Lithuanian language, which is unlikely to ever be a priority for global providers. Finally, they significantly reduce experimentation costs – an idea can be tested without long-term contracts with foreign platforms.
The new national AI guidelines for 2026–2035 declare the ambition for Lithuania to become not a technology user but a creator. The era of open models is probably the most realistic path to achieve this ambition.
Enthusiasm must be accompanied by discipline
At the same time, it would be a mistake to equate openness with security or free of charge. “Kimi K3” currently lacks a publicly released technical report or independent security assessment, and recent studies show that open-weight models can hide hard-to-detect “traps” for a relatively small amount – from biased answers to generating harmful code. Therefore, an organization implementing such a model in sensitive processes must know where the model weights came from, who modified them, and how the model’s behavior is monitored. From August, the EU AI Act will turn these issues from good practice into a legal obligation.
There is also simpler arithmetic. A cheaper token price does not necessarily mean a cheaper result. The first independent tests show that “Kimi K3” consumes significantly more computing resources to generate answers than competitors, so real costs should be evaluated based on the specific task performed, not the price list. Moreover, managing a model of this size independently requires serious infrastructure. Therefore, for most Lithuanian organizations, a more rational choice will likely be smaller open models or cloud solutions from trusted European providers.
The most important thing is not whose model it is, but who knows how to use it
The story of “Kimi K3” once again confirms a trend that has been emerging for some time: AI technology is commoditizing much faster than anyone predicted. When a top-level (frontier) AI model can be downloaded for free, the competitive advantage shifts elsewhere – to the ability to integrate the technology into processes, data, products, and decision-making.
This is the same conclusion reached by Lithuania’s position in European AI rankings: tool availability is no longer a barrier, but the ability to create value from them is still rare. The wave of open models only strengthens this trend.
Lithuania does not need to choose sides between Washington and Beijing – it needs to develop competence that allows any model to be soberly evaluated, safely adapted, and turned into real value. Ultimately, in the AI race, those who will win are not those with the most powerful model, but those who learn to use it fastest.
