Kimi K3 has already torn through the entire Silicon Valley
Four events unfolded on the same day, July 22.
Michael Kratsios, Director of the White House Office of Science and Technology Policy, publicly accused Moonshot AI's Kimi K3 on X of allegedly engaging in "massive distillation" of US models.
Greg Brockman, President of OpenAI, acknowledged in a Bloomberg interview that K3 "is a very good model, no doubt about it."
Meanwhile, nearly 200 Silicon Valley startups signed a joint letter to the Trump administration, warning that banning Chinese open-source models would cause "hundreds of companies to die instantly." Jensen Huang, CEO of Nvidia, explicitly stated in an exclusive Axios interview that US companies "absolutely should be allowed to use Chinese models."
Four voices, four directions, all pointing to the same trigger.
This trigger is Kimi K3, released by Chinese AI company Moonshot AI on July 16.
If the shockwave ignited by DeepSeek in early 2025 primarily shook Wall Street's confidence in AI infrastructure investment, the round of impact triggered by K3 targets a more fundamental issue — what should the US AI industry do when China can release model weights at a level close to the cutting edge?
Kimi K3, whose capabilities are advancing rapidly, tore apart the entire Silicon Valley and Washington in just one week.
01
Not "Just Another Chinese Model"
Let's start with Kimi K3 itself.
Featuring 2.8 trillion parameters, a Mixture of Experts (MoE) architecture with 16 out of 896 experts activated per inference, and supporting a 1 million-token context window, it is currently the world's largest open-weight model. Its full weights will be publicly released on July 27.
On Artificial Analysis's intelligence index, K3 scores 57, ranking third — second only to Anthropic's Fable 5 (60 points) and OpenAI's GPT-5.6 Sol (59 points), and surpassing Claude Opus 4.8 (56 points). In front-end coding capabilities, K3 claimed the top spot on the Arena leaderboard within 24 hours of its release, outperforming all leading US models.
Technically, K3 introduces several interesting architectural innovations.
Kimi Delta Attention (KDA) is a hybrid linear attention mechanism that achieves a 6.3x decoding speedup in million-token-level contexts. Attention Residuals (AttnRes) allows each layer to selectively retrieve information from any earlier layer, instead of uniformly accumulating information like traditional residual connections. Combined with Stable LatentMoE and quantization-aware training starting from the SFT phase, K3 delivers an approximately 2.5x overall scaling efficiency improvement over its predecessor K2.
But what truly makes the US nervous is not the benchmark scores, but the simultaneous occurrence of three factors: "performance, price, and open access."
K3's API price is $15 per million output tokens. This is considered expensive among Chinese peers — DeepSeek V4 costs only $0.87, and Zhipu's GLM-5.2 is priced at $4.4.
However, compared to US closed-source models, this price is less than one-third of Fable 5's price. A near-cutting-edge model that provides services at a price far lower than competitors, and also makes all its weights public — the combination of these three elements has made K3 the Chinese AI product that has had the greatest impact on Silicon Valley since DeepSeek.
And this did not emerge out of nowhere.
Prior to K3, Kimi's series of AI models had already quietly penetrated Silicon Valley's production chains.
Cursor, the code tool that SpaceX is acquiring for approximately $60 billion, runs its core Composer 2 on Kimi K2.5. Andy Fang, CTO of DoorDash, publicly stated that the company has already assigned "low-level tasks" to Kimi K2.6. Thinking Machines uses K2.5 to generate early post-training data for its new model Inkling.
The release of K3 is less a debut of a new model than a public confirmation of the fact that Chinese open-source AI has "already been embedded in production workflows" in Silicon Valley.
02
A Divided Silicon Valley
If the story was just about "a new Chinese model scoring high on benchmarks," it would end here.
But in the week following K3's release, the real explosive development is not technical evaluations, but the reactions it sparked in the US. To sum it up in one sentence — for the first time, Silicon Valley has had an open, fierce, and almost camp-based confrontation over "whether open-source models are a good thing or a bad thing."
The first camp is the White House and policy hawks.
Kratsios's accusations on X were phrased harshly, claiming that Moonshot "developed a sophisticated internal platform to distill US models at scale, and can rapidly switch between multiple access methods to avoid detection." The previous day, Treasury Secretary Bessent sent a similar signal, mentioning that sanctions could be considered if "watermarks" of US models are found in Chinese models. Both houses of Congress are also advancing legislation targeting unauthorized distillation. The logic of this camp is clear — there may be unwarranted improper means behind K3's capabilities, and policy and legal tools are needed to address it.
The second camp consists of US closed-source labs. OpenAI's attitude is the most thought-provoking. President Brockman acknowledged K3's strengths, but shifted the conversation to OpenAI's advantage in infrastructure investment, arguing that open-weight models are not truly "free" because large-scale deployment still requires expensive hardware. He estimated that China's overall model capabilities still lag behind the US by about 4 months.
But what better represents the sentiment of this camp is the long tweet written by Dean Ball, Head of Strategic Futures at OpenAI, the day after K3's release.
Ball claimed that open-source models are "inherently decelerationist," because they will erode the profit margins of cutting-edge labs, reduce continuous investment in AI infrastructure, and ultimately slow down the development of the most powerful models.
He even predicted that a world dominated by open-weight models would move toward "AI communism" — a future he described as a "dystopian nightmare." He also suggested that the best strategy for the Trump administration is to "create a lot of regulatory risk" for Chinese open-source models, using FUD (Fear, Uncertainty, and Doubt) to make US companies voluntarily stay away.
There is no doubt that this blatant "calculation behind ideology" immediately sparked huge controversy.
Because the third camp, US startups and open-source supporters, completely reject this view.
A new organization called the "Little Tech Association" stepped forward, with members including Proton, Replit, and Y Combinator. They organized nearly 200 companies to sign a joint letter to the White House, with one core message — if Chinese open-source models are banned, it will not be Chinese companies that die, but US entrepreneurs.
Suhail Doshi, founder of startup Particle, put it more bluntly: "Hundreds of companies will die instantly. That would be great for Anthropic — we'd all have to pay to use Anthropic's services."
The most influential voice in this camp comes from Jensen Huang, founder of NVIDIA. As the leader of the world's largest AI chip supplier, he offered a completely different judgment from the White House in his Axios interview.
Jensen Huang said Wall Street misread DeepSeek the first time, and now it is misreading K3. His logic is simple — free AI is good for chips, good for data centers, and good for hardware. Cheaper open-source models will allow more people and businesses to use AI, which will increase rather than reduce demand for computing infrastructure.
"There is no scenario where China drives US companies out," Jensen Huang said. "Zero possibility."
He also refuted the claim that open-source models "leave backdoors for China," arguing that enterprises can customize and control models in secure sandboxes. In his view, openness makes AI safer, not more dangerous — because external researchers can audit models, discover vulnerabilities, and build defenses.
"If everything becomes a single model, a single attack surface, a single point of failure, the world will be far more fragile."
03
The Debate Over "Decelerationism"
Putting aside political noise, the argument raised by Dean Ball that "open source is decelerationism" actually touches on a real industrial logic issue that deserves careful analysis.
His reasoning chain goes like this — developing cutting-edge models requires billions of dollars in investment. If a Chinese lab can provide a near-equivalent open-source alternative at extremely low cost, the profit margins of closed-source labs will be compressed. Lower profits mean less capital available for reinvestment, while the capital market will downgrade the terminal value of these companies, further constraining financing. The combination of these two effects will ultimately slow down the development of state-of-the-art models.
AI researcher Nathan Lambert acknowledged in his analysis that from a purely economic perspective, Ball's logic holds — open source does create a "deceleration" effect on cutting-edge labs at the economic level. But he also pointed out that this effect is not enough to stop OpenAI and Anthropic from becoming the most valuable companies in the world.
The market pie is growing. Even if open-source models take a portion of the share, closed-source models still have strong moats in brand trust, reliability, and service ecosystems among enterprise customers.
But the voices of opponents are equally sharp. Multiple commentators pointed out a ironic fact — OpenAI, where Ball works, and the entire modern AI field, are built on open source. The Transformer architecture came from a public paper, PyTorch is open-source software, and the research foundations and technology sharing that made OpenAI's breakthrough possible all happened before OpenAI "closed the door behind itself."
Now accusing open source of being "decelerationism" is like using open-source tools developed by others to build a skyscraper, then telling latecomers that the building doors can no longer be opened.
Some even called Ball's rhetoric "digital McCarthyism," arguing that closed-source labs are trying to use national security narratives to consolidate their market positions.
Underneath this debate is essentially a path selection problem. If large models are a new industrial foundational capability, should they be widely supplied and globally accessible like electricity and the internet? Or should they be strictly controlled by a small number of institutions like nuclear technology?
US closed-source labs see a zero-sum game — China's openness is eroding their commercial space. But from another perspective, the rapid advancement of open-source models is also doing one thing: continuously reducing the cost of AI usage, enabling more developers and enterprises that previously could not afford cutting-edge models to access these capabilities.
These two perspectives are not mutually exclusive, but they lead to completely different policy conclusions.
04
China's Models Follow Their Own Rhythm
The capital market has already given its initial reaction — in the week of K3's release, the Philadelphia Semiconductor Index fell by 12.5%, marking its largest drop in 15 months. Nvidia, AMD, and Broadcom all declined. Even China's own AI concept stocks were not spared; Zhipu fell 28% on the Hong Kong Stock Exchange, and MiniMax dropped 16%.
Jensen Huang said the market got it wrong again. But this time, what the market is panicking about is not the benchmark score of a certain model, but the fact that a path conflict has become unavoidable.
Taking a step back, the root of all current debates in Silicon Valley lies precisely in the goal chosen by the US AI industry itself — extreme accelerationism.
The logical chain of this path is clear: treat AGI as the single highest-priority climbing target. To support this goal, adopt a profit-maximizing closed-source business model, use extreme capital density to pile up extreme computing power, and ultimately achieve a capability leap that leaves the rest of the world behind. The financing scale, valuation multiples, and computing power investment of OpenAI and Anthropic are all products of this path. The "open-source deceleration" that Dean Ball is worried about essentially means that if this profit engine is weakened by open-source models, the fuel for extreme acceleration will be reduced.
But while extreme accelerationism creates competitiveness, it inevitably spawns another path.
When US labs enjoy extreme profits and extreme capital, Chinese AI companies are not in the same environment. They do not have the same level of capital density. But life finds a way — under constraints, they naturally develop their own road signs.
The evolution of the path taken by Yang Zhilin, founder of Moonshot AI, is the most vivid footnote to this path.
In October 2023, he made it very clear in a conversation at GeekPark — "Closed source is the only path to a Super APP" and "We have no plans for open source for the time being." Back then, Moonshot AI wanted to build a C-end super application through a closed-source approach.
But after the DeepSeek shockwave in 2025, Yang Zhilin made a straightforward adjustment. The trillion-parameter K2 was open-sourced, and K2.5 quickly became popular in the open-source community — its revenue 20 days after release exceeded its total revenue for all of 2025, with personal subscription orders skyrocketing by more than 80 times month-on-month. K3 continued along the open-source path to reach 2.8 trillion parameters.
This is not a random strategic swing.
In March this year, as the only Chinese independent large model founder invited to speak at Nvidia GTC 2026, Yang Zhilin presented a very clear technical roadmap in his speech "How We Scaled Kimi K2.5" — instead of blindly chasing the closed-source cutting edge by simply stacking parameters and data, Moonshot replaced all three "foundations" that the Transformer era had used for nearly a decade. The optimizer Adam (2014), the attention mechanism (2017), and residual connections (2015) — Moonshot provided three alternative solutions, all open-sourced. Elon Musk commented that it was "impressive," and Andrej Karpathy, former co-founder of OpenAI, lamented that people's understanding of the landmark Transformer paper might still be insufficient.
Four months later, K3 basically fully delivered on the roadmap in that speech — token efficiency, long context, and agent cluster collaboration were all implemented simultaneously.
At a Reddit AMA event at the end of 2025, when the Moonshot AI team was asked about the gap with US peers, they said — "Our number of GPUs is indeed less than that of US peers, but we have pushed the performance of each card to the extreme." When co-founder Zhou Xinyu was asked about his thoughts on OpenAI's heavy spending, he answered more casually — "We don't know, only Sam knows. We have our own rhythm."
"Own rhythm" — these five words may be the most important footnote to understanding the entire Chinese AI path.
It's not just Moonshot AI. The DeepSeek team built state-of-the-art models with far fewer resources. Zhipu has been continuously iterating the GLM series. DeepSeek V4 was released in April this year, GLM-5.2 went online in early July, and the WAIC conference was just held in Shanghai. This is not a story of "here comes another Chinese model" — it is a continuous, dense, multi-point concurrent rhythm of capability release. They leverage their respective strengths — underlying architecture innovation, engineering efficiency optimization, and open-source ecosystem building — to find another path approaching the cutting edge without the extreme capital density seen in the US.
OpenAI's Brockman said China lags only about 4 months behind in overall capabilities. This figure does not prove backwardness; instead, it shows that the two paths are approaching the cutting edge in their own ways.
This will not be a winner-takes-all story.
If large models are truly a new industrial foundational capability, their destiny is not to be monopolized by one company or one country, but to be used by more people. The US has taken a capital-driven, closed-source-focused path with AGI as its ultimate goal. China has developed an efficiency-driven path that uses open source as a tool to deliver capabilities more broadly. Both paths have their own advantages and costs, but together they are raising the overall level of AI.
On July 27, the full weights of K3 will be publicly released. At that time, any developer around the world can download, fine-tune, and deploy this 2.8 trillion-parameter model. That will be the moment when the game between all parties truly reaches a white-hot stage.
But a more important judgment than July 27 may be this one — from DeepSeek to Kimi, the long-term coexistence of the US and Chinese AI engines as the core driving forces of global AI is no longer a debatable prediction, but an inevitable reality that is taking place. K3 is just the latest proof.