Kimi K2.7 and Minimax M3: while the US blocks Mythos 5, China advances at an impressive speed

While the United States government ordered Anthropic to shut down Fable 5 and Mythos 5, citing national security risks, something significant and almost ironic was happening on the other side of the world: China was releasing Kimi K2.7 and Minimax M3, two new frontier models that demonstrate, once again, how American restrictions are producing the opposite effect of what was desired.

The US government's reasoning is understandable in its internal logic: limit access to the most powerful models to reduce the risk of foreign actors using them for offensive purposes. But there is a fundamental problem with this strategy: China does not need access to American models. It is building them itself, with its own resources, at rates that are surprising even the most pessimistic analysts regarding Chinese technological capability. Kimi K2.7 and Minimax M3 are the most recent proof of this trend, and ignoring them would be a strategic error of historic proportions.

Kimi is the model developed by Moonshot AI, one of the Chinese AI startups that has attracted global attention and capital in recent years. Kimi K2.7 represents the latest iteration of this family of models and has historically distinguished itself for its ability to manage very large context windows, one of the most complex challenges in Large Language Model architecture. Kimi K2.7 takes this capability to new levels, allowing for the processing of documents, code, and conversations of sizes that few models in the world can handle with comparable consistency and quality.

Internal benchmarks and community-shared tests show Kimi K2.7 to be particularly strong in tasks requiring multi-step reasoning: advanced mathematics, formal logic, and complex planning. These are areas where Chinese models had historically shown gaps compared to American competitors, but which Moonshot AI seems to have successfully addressed.

Furthermore, Kimi K2.7 integrates multimodal capabilities—text, images, and code—into a single architecture, reducing the need for complex pipelines with separate specialized models. This makes it particularly suitable for enterprise applications that require unified processing of heterogeneous content. One of the most discussed aspects is its quality-to-computational-cost ratio: Moonshot AI optimized the model to perform well even on hardware that is less extreme than that of its American competitors, a strategic choice reflecting the need to operate in a hardware ecosystem partially isolated by Nvidia chip export restrictions.

Minimax is another name well-known to those following the Chinese AI landscape. The company has stood out for a pragmatic, product-oriented approach, releasing models that quickly find application in consumer and enterprise products. Minimax M3 is its latest flagship and represents a significant leap forward compared to previous versions.

Results published by Minimax position M3 competitively against Western frontier models on standardized benchmarks such as MMLU, HumanEval, MATH, and other reference tests. This is not about absolute superiority—the frontier model landscape is still dominated by GPT-5, Claude, and Gemini Ultra in the highest categories—but about real and growing competitiveness.

Minimax M3 utilizes an optimized variant of the Mixture of Experts architecture, the same approach adopted by Mixtral and the latest generations of Gemini. The Minimax implementation introduces some proprietary innovations in the expert routing mechanism which, according to released technical papers, improve both output quality and computational efficiency during inference.

One of the most distinctive elements of Minimax M3 is its multimodal generation capability that goes beyond text and images: M3 natively integrates video and audio generation capabilities, positioning itself not just as a language model but as a unified creative platform. Added to this is aggressive pricing, significantly lower than equivalent American competitors, a strategy that is creating pressure on the margins of OpenAI, Anthropic, and Google in the competition for global enterprise customers.

To understand the scope of what is happening, it is necessary to take a step back. The speed at which China is releasing competitive frontier models is not a matter of chance, but the result of precise strategic choices made at both the governmental and industrial levels. The Chinese government has identified AI as a national strategic priority for over a decade, with five-year plans providing for investments in the hundreds of billions of yuan, creating an ecosystem where startups like Moonshot AI, Minimax, DeepSeek, and Zhipu AI operate in a context of state support that has no equivalent in the West.

The American restrictions on the export of advanced chips, instead of stopping Chinese AI development, have stimulated a wave of innovation in computational efficiency. Chinese researchers have developed training and inference techniques optimized for less powerful hardware, techniques that have also proven advantageous on a global scale. DeepSeek R1, at the beginning of 2025, was already a clear example of this.

Added to this is a massive base of technical talent, with hundreds of thousands of graduates in computer science, mathematics, and engineering fueling the sector every year, and a culture of rapid release: Chinese AI companies tend to adopt a much more aggressive approach than their American counterparts, with less focus on public communication and more focus on iteration speed.

To have a clear view of where we are, it is useful to compare the state of the art of the two ecosystems. The USA maintains an advantage in the most advanced frontier models—GPT-5, Claude Opus, and Gemini Ultra remain the reference benchmarks—but the gap is narrowing at a speed that few had predicted. China, on the other hand, has an advantage in terms of iteration speed, releasing frequent updates with less ceremony but more substance.

Hardware restrictions have turned what was a disadvantage into a competitive advantage: Chinese models tend to be more efficient per unit of performance than American competitors, a characteristic that is increasingly relevant as the cost of inference becomes a critical competitive factor. Furthermore, in China, AI models have a much faster penetration in consumer applications than in the West: WeChat, Alipay, Baidu, and dozens of other platforms have integrated advanced AI capabilities, creating a virtuous feedback loop based on real usage data.

Returning to the starting point: the American government blocked Mythos 5 citing national security risks related to the possibility of foreign actors exploiting its advanced software system analysis capabilities. But Kimi K2.7 and Minimax M3, released almost simultaneously, have comparable capabilities in the same areas. This creates an obvious strategic paradox: American restrictions do not reduce the global availability of advanced AI, but rather shift the center of gravity of development towards China.

A foreign actor wishing to access advanced AI capabilities for offensive purposes does not need to bypass Anthropic's protections on Mythos 5: they can simply use Kimi K2.7, or Minimax M3, or DeepSeek, or any other Chinese model not subject to American restrictions. Restrictions, in this sense, primarily harm American companies and researchers and their allies, without removing the risk they claim to mitigate.

If the strategy of unilateral blocking is structurally ineffective, what is the alternative? Various AI policy experts are converging on a set of complementary approaches: investing in leadership instead of blocking, cooperating internationally on safety standards, developing governmental capacities for technical model evaluation, and strengthening the defenses of critical systems instead of merely blocking the tools that might identify their vulnerabilities.

For Europe, the Fable 5/Mythos 5 affair on one hand and Kimi K2.7/Minimax M3 on the other contains a clear message: dependence on American AI is a strategic risk, exactly as dependence on Russian energy was. If the American government can block its own AI models with an executive order, it can do the same for European customers. If China advances rapidly with its own models, Europe risks finding itself squeezed between two AI ecosystems it does not control.

The European digital sovereignty project, already underway with the AI Act and initiatives like GAIA-X, must accelerate significantly. It is not about technological protectionism, but about ensuring Europe the ability to control its own strategic infrastructures in the age of artificial intelligence.

The launch of Kimi K2.7 and Minimax M3 simultaneously with the blocking of Mythos 5 is not a narrative coincidence; it is a perfect snapshot of the state of the AI world in 2026. A world where technology advances on multiple fronts simultaneously, where no single country has a monopoly on innovation, and where unilateral control strategies show their structural limits in real-time. The future of AI is multipolar: there will not be an absolute winner, but different ecosystems with different strengths that will continuously influence and challenge each other.