Kimi K3 Ignites Global AI Debate with 2.8T Parameters and Open-Source Model

On July 22, Chinese AI company Moonshot AI released Kimi K3, causing ripples across Silicon Valley. K3 boasts 2.8 trillion parameters, making it the world’s largest open-weight model, ranking third in multiple benchmarks, yet costing less than one-third of the price of U.S. proprietary models. Its release sparked four distinct reactions: the White House accused it of “distilling” U.S. models, OpenAI’s CEO acknowledged its technical prowess, nearly 200 Silicon Valley startups jointly warned against banning Chinese open-source models, and NVIDIA CEO Jensen Huang explicitly supported openness. Chinese AI is advancing toward the frontier through an efficient, open-source-driven approach, achieving this without extreme capital density. A dual-engine AI landscape between China and the U.S. is taking shape.

Article author and source: GeekPark

On July 22, four events occurred on the same day.

Office of Science and Technology Policy Director Michael Kratsios publicly accused Moonshot AI's Kimi K3 of "large-scale distillation" of U.S. models on X.

OpenAI President Greg Brockman, in an interview with Bloomberg, acknowledged that K3 is “a very good model, no doubt.”

Meanwhile, nearly 200 Silicon Valley startups jointly wrote a letter to the Trump administration, warning that banning Chinese open-source models would cause "hundreds of companies to die instantly." Nvidia CEO Jensen Huang, in an exclusive interview with Axios, clearly stated that U.S. companies "should absolutely be allowed to use Chinese models."

Four sounds, four directions, pointing to the same trigger.

The trigger was Kimi K3, released by the Chinese AI company Moonshot on July 16.

If the shockwave triggered by DeepSeek in early 2025 primarily challenged Wall Street’s faith in AI infrastructure investment, then the wave ignited by K3 targets a more fundamental question—when China can open model weights at near-cutting-edge levels, what should America’s AI industry do?

The increasingly powerful Kimi K3 tore through Silicon Valley and Washington in just one week.

First, let’s talk about Kimi K3 itself.

2.8 trillion parameters, MoE (Mixture of Experts) architecture, with 16 experts activated out of 896, supporting a context window of 1 million tokens. This is currently the largest open-weight model globally, with full weights to be publicly released on July 27.

On the Artificial Analysis intelligent index, K3 scores 57 and ranks third—behind Anthropic’s Fable 5 (60) and OpenAI’s GPT-5.6 Sol (59), but ahead of Claude Opus 4.8 (56). In frontend coding ability, K3 rose to number one on the Arena leaderboard within 24 hours of release, surpassing all leading U.S. models.

Technically, K3 introduces several interesting architectural innovations.

Kimi Delta Attention (KDA) is a hybrid linear attention mechanism that achieves 6.3x faster decoding over contexts of millions of tokens. Attention Residuals (AttnRes) enable each layer to selectively retrieve information from any earlier layer, rather than uniformly accumulating it as in traditional residual connections. Combined with Stable LatentMoE and quantization-aware training starting from the SFT stage, K3 delivers approximately 2.5x overall scaling efficiency improvement over its predecessor, K2.

But what truly concerns the United States is not the ranking numbers, but the simultaneous occurrence of performance, price, and openness.

The API price for K3 is $15 per million output tokens. This is expensive compared to Chinese peers—DeepSeek V4 costs only $0.87, and Zhipu’s GLM-5.2 is $4.40.

But at this price, it’s less than one-third the cost of Fable 5 compared to closed-source models in the U.S. Offering a model that approaches state-of-the-art performance at a fraction of competitors’ prices—while fully open-sourcing its weights—these three factors combined make K3 the Chinese AI product with the biggest impact on Silicon Valley since DeepSeek.

And this didn't come out of nowhere.

Before K3, Kimi's series of AI models had already quietly infiltrated Silicon Valley's production chain.

SpaceX is acquiring the code tool Cursor for approximately $60 billion, and the core of its Composer 2 runs on Kimi K2.5. DoorDash’s CTO Andy Fang publicly stated that the company has delegated “low-level tasks” to Kimi K2.6. Thinking Machines is using 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 and more a public acknowledgment that Chinese open-source AI is already embedded in production workflows in Silicon Valley.

If it's just "a new Chinese model scores very high," the story ends here.

But a week after K3's release, what truly exploded wasn't the technical reviews, but the reaction it sparked in the United States—in one sentence, Silicon Valley experienced its first open, intense, and nearly factionalized debate over whether open-source models are a good or bad thing.

The first camp is the White House and policy hawks.

Kratsios’s allegations on X were quite severe, claiming that Moonshot “developed a sophisticated internal platform to systematically distill U.S. models and rapidly switch between multiple access methods to avoid detection.” The previous day, Treasury Secretary Bessent signaled a similar stance, suggesting that sanctions might be considered if “watermarks” from U.S. models were found in Chinese models. Both houses of Congress are also advancing legislation targeting unauthorized distillation. The logic of this faction is clear—the capabilities behind K3 may stem from unfounded, improper means, and must be addressed using policy and legal tools.

The second camp consists of closed-source labs in the United States. OpenAI’s stance is particularly intriguing: President Brockman acknowledged K3’s capabilities but shifted the focus to OpenAI’s advantages in infrastructure investment, arguing that open-weight models are not truly “free,” as large-scale deployment still requires expensive hardware. He estimated that China still lags behind the United States by approximately four months in overall model capability.

But更能代表这一派情绪的，是 OpenAI 战略未来主管 Dean Ball 在 K3 发布次日写的那条长推文。

Ball called open-source models "inherently decelerationist," as they erode the profit margins of leading labs, reduce ongoing investment in AI infrastructure, and ultimately slow the development of the most advanced models.

He even predicted that a world dominated by open-weight models would lead to "AI communism"—a future he described as "like a dystopian nightmare." He also advised the Trump administration that its best strategy would be to create substantial regulatory risks for China’s open-source models, using FUD (fear, uncertainty, doubt) to encourage American companies to steer clear on their own.

Undoubtedly, this blatant "calculation behind the ideology" immediately sparked significant controversy.

Because the third camp—American startups and open-source advocates—completely reject it.

A new organization called the Little Tech Association, comprising Proton, Replit, and Y Combinator, has rallied nearly 200 companies to sign a joint letter to the White House with one core message: if Chinese open-source models are banned, it won’t be Chinese companies that die—it will be American entrepreneurs.

Suhail Doshi, founder of startup Particle, put it more bluntly: “Hundreds of companies will die overnight. This is fantastic for Anthropic—we’ll all have to spend money buying Anthropic’s services.”

The most prominent voice in this camp comes from Jensen Huang, founder of NVIDIA, who, as the leader of the world’s largest AI chip supplier, offered a judgment in his Axios interview that directly contradicts that of the White House.

Huang Renxun said Wall Street misread DeepSeek for the first time, and now it’s misreading K3. His logic is simple—free AI is good for chips, good for data centers, good for hardware. Cheaper open-source models will enable more individuals and businesses to adopt AI, increasing demand for computing infrastructure, not reducing it.

"There is no scenario where China drives out U.S. companies," said Jensen Huang. "Zero possibility."

He also refuted the claim that open-source models leave backdoors for China, arguing that companies can customize and control models within secure sandboxes. In his view, openness makes AI safer, not more dangerous—because external researchers can review the models, identify vulnerabilities, and build defenses.

If everything becomes a single model, a single attack surface, and a single point of failure, the world would be much more vulnerable.

Setting aside political noise, Dean Ball's argument that "open source is slowdownism" touches on a genuine industrial logic issue worthy of careful analysis.

His reasoning chain is this—developing state-of-the-art models often requires billions of dollars in investment. If a Chinese lab can provide an open-source alternative with near-equivalent performance at a fraction of the cost, the profit margins of closed-source labs will be squeezed. Reduced profits mean less capital available for reinvestment, while capital markets will lower the terminal valuations of these companies, further dampening fundraising. The combined effect will ultimately slow the pace of development for the most advanced models.

AI researcher Nathan Lambert acknowledges in his analysis that, from a purely economic perspective, Ball’s logic holds—open source does indeed create a “slowing” effect on leading labs. However, he also points out that this effect is not sufficient to prevent OpenAI and Anthropic from becoming the world’s most valuable companies.

The market pie is growing, and even though open-source models have captured a portion of the share, proprietary models still maintain a strong moat in terms of brand trust, reliability, and service ecosystems for enterprise customers.

But the critics’ voices are equally sharp. Several commentators have pointed out a ironic fact—that OpenAI, where Ball is based, and the entire modern AI field, are themselves built on open-source foundations. The Transformer architecture is from an open paper, PyTorch is open-source software, and the research foundations and technological sharing that enabled OpenAI’s breakthroughs all occurred before OpenAI “closed the door behind it.”

Accusing open source of being "decelerationism" today is like building a building using someone else's open-source tools, then telling newcomers that the door can't be opened anymore.

Some have labeled Ball's stance as "digital McCarthyism," accusing closed-source labs of attempting to leverage national security narratives to solidify their market position.

At its core, this debate is a question of path selection. If large models are a new industrial foundational capability, should they be widely distributed and globally shared, like electricity and the internet? Or should they be strictly controlled by a few institutions, like nuclear technology?

Closed-source labs in the U.S. 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 lowering the cost of AI access, enabling developers and businesses that could previously not afford cutting-edge models to gain capability.

These two perspectives are not mutually exclusive, but they lead to vastly different policy conclusions.

The capital markets have already delivered their initial response—during the week of K3's release, the Philadelphia Semiconductor Index fell 12.5%, marking its largest drop in 15 months. Nvidia, AMD, and Broadcom all declined sharply. Even China’s own AI-themed stocks were not spared, with Zhipu falling 28% on the Hong Kong Stock Exchange and MiniMax dropping 16%.

Jensen Huang said the market got it wrong again. But this time, what the market is panicking about isn't a model's benchmark score—it's an unavoidable path conflict.

Looking at it from a step back, the root of all the current debates in Silicon Valley stems precisely from the goal that the U.S. AI industry itself chose—extreme accelerationism.

The logic of this path is clear—position AGI as the singular, highest-priority goal, support it with a profit-maximizing closed-source business model, and deploy extreme capital density to build extreme computing power, ultimately achieving a leap in capability that leaves the rest of the world behind. The funding scales, valuation multiples, and compute investments of OpenAI and Anthropic are all products of this path. What Dean Ball fears as “open-source slowdown” is essentially this: if this profit engine is weakened by open-source models, the fuel for extreme acceleration will diminish.

But extreme accelerationism, while creating competitiveness, will inevitably give rise to another path.

When American labs enjoy extreme profits and extreme capital, Chinese AI companies do not operate in the same environment. They lack the same level of capital density. But life finds a way—under constraints, it naturally grows its own signposts.

The personal journey of Yang Zhilin, founder of Moonshot AI, is the most vivid illustration of this path.

In October 2023, he was very clear in his conversation at Geek Park: “Closed-source is the only path to a Super APP,” and “We have no plans to open-source for now.” At that time, Moonshot AI aimed to build a consumer-facing super app using a closed-source approach.

But after the DeepSeek shockwave in 2025, Yang Zhilin made a decisive adjustment: the trillion-parameter K2 was open-sourced, and K2.5 quickly gained popularity in the open-source community—its revenue in just 20 days after launch exceeded the entire year’s revenue of 2025, with individual subscription orders surging more than 80-fold month-over-month. K3 continued along the open-source path, reaching 2.8 trillion parameters.

This is not a casual strategic shift.

In March this year, Yang Zhilin, the only Chinese independent large model founder invited to speak at NVIDIA’s GTC 2026, presented a clear technical roadmap in his talk “How We Scaled Kimi K2.5”—instead of blindly chasing closed-source frontiers by stacking parameters and data, he replaced all three foundational components of the Transformer era, which had been in use for nearly a decade. The optimizer Adam (2014), the attention mechanism (2017), and residual connections (2015) were each substituted with new alternatives, all open-sourced by Moonshot AI. Musk called it “impressive,” and former OpenAI co-founder Karpathy remarked that our understanding of the seminal Transformer paper may still be insufficient.

Four months later, K3 delivered on nearly the entire roadmap outlined in that talk—token efficiency, long context, and agent cluster collaboration—all three lines were implemented simultaneously.

During a Reddit AMA at the end of 2025, when asked about the gap with their American counterparts, the Moonshot team replied, “We indeed have fewer GPUs than our American peers, but we’ve squeezed maximum performance out of every single card.” When questioned about OpenAI’s spending, co-founder Zhou Xinyu responded more casually: “We don’t know—only Sam knows. We have our own pace.”

"Your own pace"—these five characters may be the most important footnote to understanding China’s entire AI trajectory.

It’s not just Moonshot AI. The DeepSeek team built a model of comparable performance with fewer resources. Zhipu has been continuously iterating on the GLM series. DeepSeek V4 was released in April this year, GLM-5.2 went live in early July, and the WAIC conference was just held in Shanghai. This isn’t a story of “another Chinese model has arrived”—it’s a continuous, intense, and multi-pronged release rhythm of capabilities. Each team leverages its own strengths—innovation in underlying architecture, optimization of engineering efficiency, and building open-source ecosystems—to forge an alternative path toward the cutting edge, without the extreme capital density seen in the United States.

Brockman from OpenAI said China is only about four months behind in overall capability. This figure illustrates not a gap, but rather two paths simultaneously approaching the frontier in their own ways.

This won't be a winner-takes-all story.

If large models truly represent a new industrial foundational capability, their destiny is not to be monopolized by a single company or nation, but to be widely adopted. The United States has pursued a path driven by capital, primarily closed-source, with AGI as the ultimate goal. China has forged a path driven by efficiency, using open-source as a tool, with the logic of delivering broader capabilities. Both paths have their own advantages and costs, but together they are raising the overall tide of AI.

On July 27, the full weights of K3 will be publicly released, allowing any developer worldwide to download, fine-tune, and deploy this 2.8 trillion parameter model. That will be the moment when competition truly intensifies.

But perhaps a more important judgment than July 27 is this—that the long-term coexistence of DeepSeek and Kimi as the core driving forces of global AI is no longer a debatable prediction, but an inevitable reality already unfolding. K3 is merely the latest proof.