Moonshot AI, founded on March 1, 2023, coinciding with the 50th anniversary of Pink Floyd's iconic album "The Dark Side of the Moon," which is a favorite of founder Yan Zhiling, has evolved from a small IT office to a company poised for its final pre-IPO round in Hong Kong, with media reports suggesting a valuation could reach $50 billion.

ForkLog has explored the Chinese model's phenomenon, the advantages of open weights, and the reasons why the rules of the AI race are changing faster than policymakers can respond.

The Man from China

Born in 1992 in Shantou, Guangdong province, Yan Zhiling studied at Tsinghua University in Beijing and later at Carnegie Mellon University in the United States, where he earned his PhD in four years. During his final years, Yan interned at Google Brain and Meta, following a path that has recently been a direct route for Chinese students into the American tech sector, often leading to permanent relocation.

However, Yan chose to return to China, where he worked on Huawei's PanGu model. He was later invited to the Beijing Academy of Artificial Intelligence to develop the Wu Dao project. This trajectory is common among key figures in China's AI sector in recent years, many of whom returned home after studying abroad.

Yan Zhiling (right). Photo: Cheyenne Zhao.

Yan's PhD advisor, Ruslan Salakhutdinov, praised him as "absolutely brilliant" after the launch of Kimi K3, noting that he was sought after by major tech companies even before his defense.

This narrative—of a talent choosing to return home—can easily be turned into a propaganda story. However, the co-founders of Moonshot AI are Yan's classmates from Tsinghua University. Yan is building a company based on his preferences with trusted individuals, and the macroeconomic significance of his biography becomes apparent only when the startup reaches a scale where its actions can be interpreted as part of a national strategy.

Kimi vs. DeepSeek

Forecasts for Moonshot AI suggest that the venture capital market has lost its sense of proportion. By the end of 2025, the company was valued at $4.3 billion. By February 2026, that number had risen to $10 billion, followed by $18 billion in March, over $20 billion in May, and $35 billion by the end of July. The annual recurring revenue (ARR) doubled from March to April to $200 million, and by June, it reached $300 million.

The next target is $50 billion; this is the valuation Moonshot AI aims for in its final round before going public in Hong Kong.

Behind these figures is a company of around 300 people that was hardly mentioned in market reviews outside Asia a year ago. Despite the typical venture capital hype (from Alibaba, Tencent, Meituan) ahead of the listing, the company does indeed have verified revenue.

More than 70% of Moonshot AI's revenue comes from direct sales of API access to corporate clients and third-party developers, while the remainder consists of international paid subscriptions to Kimi.

Kimi K3 pricing plans. Source: Kimi.

Just 48 hours after the release of Kimi K3, the management had to suspend new user connections due to a shortage of computing power. This event sparked mixed reactions. On one hand, it highlighted a disconnect between the rapid growth on paper and physical limitations. The $50 billion valuation suggests that Moonshot AI can meet demand at the scale of a major market player. On the other hand, analysts noted the strong demand, confirming that the product has exceeded market expectations.

Despite Moonshot AI's success, the impact of the DeepSeek-R1 launch in January 2025 was significantly greater. Kimi K3's price gap with top models is noticeably smaller than R1's at launch, indicating a weaker price advantage.

DeepSeek-R1 caused a price shock: it offered reasoning capabilities comparable to OpenAI's o1 but was sold via API at 10-30 times lower prices—around $0.55 per million input tokens compared to $15 for o1.

This aggressive underpricing forced a complete overhaul of the inference economy in the industry.

Under current conditions, Kimi K3 lacks such an advantage: it costs $3 per million input tokens and $15 for output. Western LLMs of similar caliber (Claude Sonnet 5) are priced similarly—around $2.5-$3 for input and $15 for output. Compared to the most expensive models, such as Claude Fable 5, the Chinese alternative is about three times cheaper.

Source: Emergent.

The success of this latest Chinese model is not solely based on funding or architecture, but also on specific choices made by the founders a year prior to K3's release.

Openness is Not Altruism

Kimi was initially designed as a closed model: training details were not disclosed, and the weights remained the company's property—a standard approach for any proprietary AI product. The pivot occurred after DeepSeek garnered global attention with systems developed at relatively low costs and made them publicly available. Moonshot AI responded quickly, modifying Kimi's architecture and moving towards publishing models with open weights.

On July 11, 2025, the company released Kimi K2—the first Chinese open model with a trillion parameters. It surpassed DeepSeek-R1 in popularity, which had previously been considered the benchmark for the entire category. The release of K3, the largest LLM to date (2.8 trillion parameters), took place on July 16, 2026. The open weights were published on July 27.

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.

This is a 17-place jump from Kimi-k2.6 (#18 -> #1).

In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics,… https://t.co/YDN3BufGkC pic.twitter.com/Oa6teaQnWp

— Arena.ai (@arena) July 16, 2026

The model ranked first in an independent evaluation by the Frontend Code Arena, scoring 1679 points compared to 1631 for Claude Fable 5 and 1618 for GPT-5.6-Sol. According to the Artificial Analysis Intelligence Index, K3 ranks fourth, surpassing Claude Opus 4.8.

However, not everything is smooth for the Chinese flagship: independent testing showed an increase in hallucination rates to 51%, compared to 39% for the previous K2.6 version. The model responds more accurately on average but also produces erroneous answers more frequently.

Due to its volume of 2.8 trillion parameters, launching the model locally requires a substantial amount of RAM:

  • quantization with virtually no quality loss (4-bit) requires about 1.6 TB;
  • dynamic hard quantization (1-bit/2-bit) compresses the model weight to around 700 GB.

Theoretically, launching a severely stripped-down version without server hardware is possible if several top Mac Studios (each with a limit of 512 GB of combined memory) are networked together. However, standard network cables cannot replace the ultra-fast data center buses, and there is currently no official Apple MLX support for the K3 model.

A similar situation applies to PCs: attempting to run K3 on a machine with four RTX 5090-level GPUs and DDR5 RAM will result in generation speeds dropping to utterly unworkable levels—taking several minutes per token.

Hardware configurations for local Kimi K3 launch. Source: Kingy AI.

The openness of Moonshot AI is a response to competitive pressure on two fronts. Firstly, it is a way to attract developers and maintain a presence in the global market where closed American models are physically unavailable within China. Secondly—and more importantly—open weights create a problem for Washington that is difficult to address legislatively. A model with open weights, once published online, is immediately copied to thousands of servers worldwide, making it technically challenging to prohibit its use within the country.

The U.S. administration is already considering restrictions on advanced Chinese models within its borders. The release of K3 has catalyzed this discussion. While no formal policy or specific mechanism has been announced, the mere framing of the issue indicates that Moonshot AI's openness is perceived as a strategic concern.

In response to such rhetoric, Nvidia, Meta, Microsoft, and 22 other companies issued a letter urging U.S. authorities not to impose restrictions on AI models with open weights. The authors warn that bans will not strengthen U.S. leadership but rather shift the market towards a few closed developers. They compared the situation to the open-source movement. According to the signatories, American leadership in AI will depend not on a single cutting-edge closed model but on the country's ability to create an open ecosystem across various sectors.

After K3's release, American officials and Anthropic publicly accused the company of distilling American models—essentially training Kimi on outputs from foreign systems and violating usage conditions. The Chinese startup rejected these allegations.

Additionally, the letter's authors defended distillation, labeling it a legitimate practice. They suggested that contentious cases should be resolved through specific legal and commercial frameworks rather than through an outright ban.

Competitive pressure that forced Moonshot AI to pivot its model architecture also acts at the hardware level—where sanctions have played a complex role.

Sanctions as a Catalyst for Progress

The goal of Washington's export restrictions was clearly stated: to deny China access to the most advanced AI chips and manufacturing equipment to slow the development of systems with potential military and intelligence applications.

During trade negotiations in August 2025, Beijing requested a relaxation of restrictions specifically on high-bandwidth memory supplies rather than chips and lithography equipment in general.

There are workarounds: some Chinese companies rent computing power abroad or purchase processors through foreign subsidiaries. None of these options resolves the problem systematically, but each alleviates some pressure.

DeepSeek R1 demonstrated a way forward. The Chinese open model can compete with proprietary systems from American labs. The belief that creating a top-tier LLM requires colossal investments and computing power on par with the market's largest players has been challenged.

The gap continues to narrow, but not through technological superiority—rather through a series of incremental solutions. For example, engineering routines: hard context caching (with K3 offering a 90% discount on repeat tokens), hybrid attention architectures, and optimization of token generation reasoning.

Rules of the Game Are Changing Faster

Formally, Moonshot AI remains a private company under the founder's control: Yan retains 51.83% of the shares. Among the investors are structures affiliated with the state, including the National Social Security Fund of China—one of the country's largest institutional investors managing pension reserves. The presence of such players in the AI startup's capital is not a trivial detail but an indicator of the company's significance to Beijing.

As it prepares for listing in Hong Kong, Moonshot AI's management dismantled its offshore Variable Interest Entity (VIE) structure—a mechanism that Chinese tech companies have used for decades to attract foreign capital while circumventing restrictions on direct ownership. The abandonment of this structure is a direct result of rules adopted by the China Securities Regulatory Commission in March 2023 regarding overseas listings.

From the perspective of Chinese law, this means the company is no longer trying to circumvent foreign ownership restrictions through legal loopholes. Instead, it officially acknowledges the full jurisdiction of Chinese authorities over its corporate structure, data, and algorithms.

American media have also raised the issue of the potential military applications of Kimi technologies. There are currently no confirmed instances of such uses—this is more of a narrative circulating amidst overall wariness regarding China's AI sector.

None of the individual facts prove that Moonshot AI is a state project in a private guise. However, the combination of details paints a picture of a business that is difficult to view outside of its national context.

This is not yet a victory over American sanctions. The gap in capabilities has narrowed but not disappeared: China is closing it through architectural workarounds rather than technological superiority. However, it is also not a failure of export control—restrictions on memory and computing still compel Chinese labs to seek unconventional solutions instead of simply ramping up capacity.

The rules of the game are changing faster than policymakers can react. Open Chinese models, backed by state funds and difficult for Washington to ban physically, are already being utilized by developers worldwide— including within American companies.