Kimi K3: A Chinese open-source model that beat GPT and Claude, then got targeted by the White House

Here are three pieces of information to give you a sense of the situation.

First, a Chinese open-source large language model has reached number one globally on public benchmarks for front-end programming, outperforming both GPT and Claude.

Second, just days after its release, the White House and the Treasury Department stepped forward to accuse it of stealing technology from American companies.

Third, the valuation this company is currently negotiating is $50 billion.

You should be very familiar with this model and company: it's Moonshot AI's new model, Kimi K3, which is preparing for an IPO in Hong Kong.

So today, let's discuss: where exactly does K3 excel, why is the US so nervous, and most importantly—is this $50 billion valuation value or a bubble?

Let's look at the hard metrics first.

K3 is a 2.8 trillion parameter MoE (Mixture of Experts) model with 896 experts, activating only 16 at any given time. You can think of it as having a huge brain but only calling upon the most relevant small portion each time, balancing capability and efficiency. It natively supports image understanding, with a context window of up to one million tokens. On the day of its release, it was the largest open-source model by parameter count in the world. But large parameters are merely an entry ticket. What truly matters are its comparative results against top-tier closed-source models on public benchmarks.

On the Artificial Analysis comprehensive intelligence leaderboard, K3 scored 57 points, ranking third globally, behind only two closed-source flagships: Claude Fable 5 and GPT-5.6 Sol. This marks the first time an open-source model has broken into the top three. In the front-end development category, it directly took the global number one spot.

However, in agent-based task evaluations, K3 scored 1668 points. While it beat Zhipu's GLM, it hasn't yet caught up to Claude Fable 5's 1760 or GPT-5.6's 1748.

Therefore, the accurate statement is: K3 has not yet comprehensively surpassed overseas closed-source models, but it has genuinely entered the global top tier. Moreover, it has established its own advantages in areas such as front-end development, code generation, and long-context knowledge work.

So why has an open-source model caused such a stir? The key isn't just K3 itself, but rather that the generational gap curve between open-source and closed-source models is being rapidly flattened.

In 2023, when GPT-4 was released, the leading open-source models were still at the level of GPT-3.5, lagging by a full generation—about 12 to 18 months.

By 2024, models like Llama 3 and Qwen 2.5 had caught up to the early levels of GPT-4, narrowing the gap to six to nine months.

By 2025 and 2026, this gap has been compressed to under three months, and single-point overtaking is beginning to occur.

K3 is the landmark event of this "single-point overtaking." For the first time, an open-source model has achieved global number one in high-value core production scenarios like front-end programming.

What does this mean? It means corporate procurement logic may change. Previously, core business operations relied solely on closed-source APIs. Now, private deployment of open-source models has become a viable alternative, with costs differing by an order of magnitude. Furthermore, open-source attracts global developers, potentially accelerating iteration speeds even faster than closed-source models. This is what people refer to as the dual-wheel drive of models and domestic computing power.

Hearing this, you might already be feeling pumped up. But I must pour a bucket of cold water on you, which many commentators won't tell you.

K3 did not replicate the most crucial aspect of DeepSeek: extreme low cost. According to Artificial Analysis calculations, the average cost for K3 to complete a weighted evaluation task is approximately $0.94. While this is significantly cheaper than Claude Fable 5, guess how much DeepSeek V4 Pro costs? Approximately $0.04—a difference of more than twenty times.

Just days after K3's release, Kratsios, Director of the White House Office of Science and Technology Policy, posted on social media stating they have information indicating that Moonshot AI distilled Anthropic's Fable model to develop K3. Distillation, simply put, involves using the output of a stronger large model to train one's own model. Shortly after, Bessent stated he would sanction K3, even claiming that Moonshot AI obtained banned Nvidia GB300 chips through Thailand.

But what is the reality?

First, Moonshot AI explicitly denied these allegations.

Second, numerous AI researchers publicly questioned the timeline inconsistencies. The Fable model was only made publicly available on July 1st, while K3 was released around July 15th—only about ten days apart. Relying on large-scale distillation to train a 2.8 trillion parameter model within this timeframe is widely considered insufficient by professionals. Moreover, distillation alone cannot explain K3's strong performance.

Third, while Anthropic indeed reported detecting approximately 3.4 million interactions attributed to Moonshot AI, they did not publicly equate these interactions directly with K3's training. Additionally, there is a legal controversy: whether model outputs are protected by copyright remains unsettled within the industry, and distillation is a widely used conventional technique.

Thus, the current status of this matter is: the accusations are fierce, but there is no public chain of evidence.

We do not take sides. For investment purposes, we only need to know one thing: this unresolved political risk has now become a real variable in Moonshot AI's valuation model. Especially regarding their overseas B2B clients, who cannot possibly ignore political factors during procurement.

Alright, back to the core question. Is $50 billion too expensive?

Using the simplest Price-to-Sales (PS) ratio calculation: Moonshot AI's expected ARR for 2026 is approximately $300 million. Dividing the $50 billion valuation by this gives a PS ratio of roughly 167x. Using the same algorithm, Zhipu is around 70x, and MiniMax is slightly over 20x.

Do you see it clearly? Even among a group of AI companies that are already expensive, Moonshot AI's multiple is in a league of its own. This price already prices in three things in advance: technological breakthroughs, rapid ARR growth, and the window for the Hong Kong IPO.

So why dare they ask for such a high price? First, K3's capabilities are the highest among the three. Its scaling efficiency improved by approximately 2.5 times compared to the previous generation K2. API revenue accounts for over 70%, and overseas revenue reportedly quadrupled. These numbers are not piled up by C-end subscriptions.

But being expensive comes with problems:

First, the applicable space for cost-effectiveness is narrow. As mentioned, it is expensive. For tasks like customer service, summarization, copywriting, and simple coding, GLM, MiniMax, or even smaller models suffice, and they are faster and cheaper. K3 is only worthwhile in complex scenarios where task value is high and the cost of failure is significant.

Second, the deployment threshold is high. The official recommendation is to use super-nodes with more than 64 cards. Rough estimates suggest that local deployment alone requires a conservative hardware investment of 30 million RMB. Furthermore, within 48 hours of K3 going online, request volumes hit cluster limits, forcing the company to temporarily suspend new C-end subscriptions. Supply elasticity clearly cannot keep up.

Third, reliability is questionable. Their own B-end head admitted that hallucinations cannot be completely eradicated. Therefore, in serious scenarios like law, auditing, and academia, manual review is still required. This directly limits the speed of its adoption in high-value scenarios.

Fourth, there is the aforementioned overseas policy risk, which I will not repeat.

So, to summarize: K3's technical weight is undeniable. It has pushed domestic open-source models into the global top tier for the first time. However, a powerful model and a reasonable valuation are two different things.

The ceiling of this company depends on whether K3 can continue to dominate those high-value complex tasks; its floor depends on ARR, customer retention, computing power costs, and policy risks.

If its lead can withstand more than two iteration cycles, pushing ARR from $300 million to $700-$800 million, or even $1 billion, then the PS ratio would drop from 167x to 60-70x, making $50 billion actually not expensive. But if Zhipu and MiniMax catch up quickly, K3's advantage would remain only the 30-40 day release time gap, and this multiple would not be supported by earnings.

So, in one sentence: Moonshot AI currently has a high ceiling but low certainty. The model is genuinely impressive, but whether it is worth $50 billion, I have a big question mark.