The real focus isn't the model score, but Yang Zhilin’s chosen path: open weights + long context + agent clusters—a route distinctly different from OpenAI’s.

Author and source: 0x9999in1, ME News



TL;DR

• On July 16, 2026, Moonshot AI released Kimi K3, featuring 2.8 trillion parameters and a 1 million token context window. Independent evaluation by Artificial Analysis yielded a composite score of 57.1, ranking it 4th globally, surpassing Claude Opus 4.8.
• Founder Yang ZhiLin, 34, holds a bachelor’s degree from Tsinghua University and a Ph.D. from Carnegie Mellon University. He is the lead author of Transformer-XL and XLNet, among the rare individuals who turn their own research papers into products.
• In May 2026, Moon Shadow completed a new funding round of approximately $2 billion, achieving a valuation exceeding $20 billion—roughly quadrupling its valuation within three months.
• On the day K3 was launched, the Nasdaq fell more than 1.4%, and NVIDIA also weakened—marking the first time a Chinese startup’s product launch genuinely influenced U.S. market sentiment.
• But K3 is no longer "extremely cheap": at $3/$15 per million tokens, it is about three times the price of the previous K2.6, marking a turning point for the era of "Chinese AI at rock-bottom prices."
• The hallucination rate of K3 was independently evaluated at as high as 51%, the output is verbose, and the open weights have been delayed until July 27— the story is far from perfect.
• The real focus isn't the model score, but Yang Zhilin’s chosen path: open weights + long context + agent clusters—a route distinctly different from OpenAI’s.

A drummer, why could he create the fourth-largest model?

First, the conclusion: Yang Zhilin is not just lucky.

He is one of the very few who can turn theory into a finished product.

Transformer-XL and XLNet, two papers with the first author. These are not ordinary papers—they are among the foundational works underlying today's large models for processing long sequences.

Many people write papers. Few turn papers into companies. And worldwide, you can count on one hand the number of companies that have made it into the global top four.

Born in 1992 in Shantou, Guangdong. Earned a bachelor’s degree in Computer Science from Tsinghua University, where he played drums and wrote songs for the campus rock band Splay while coding. Completed his Ph.D. at Carnegie Mellon University in four years. Has worked at Google Brain and Meta AI, with collaborators including Bengio and LeCun.

Co-founded his first AI company at age 24. Founded Moonshot AI in March 2023.

This resume is a rare specimen even in Silicon Valley, and nearly unique in China.

Why emphasize this?

On the path of large models, "I have read all the relevant papers" and "these papers were written by me" represent two entirely different levels of potential. The former is a learning curve; the latter is first-mover intuition. Yang Zhilin belongs to the latter.

Transformer-XL addresses the problem of "forgetting long sequences." XLNet advances beyond BERT’s pretraining paradigm. On this extended trajectory lies Kimi’s most distinctive product label today: long context.

From the early Kimi AI assistant’s 200,000 tokens, to the later 2 million tokens, and now directly offering 1M tokens (approximately 1 million tokens) with K3.

This is not a marketing topic; it is a natural extension of his academic career.

What are the strengths and weaknesses of Kimi K3?

First, let's look at the hard data.

Kimi K3 was released on July 16, 2026. It features 2.8 trillion total parameters, a sparse MoE architecture activating 16 experts per token out of 896 total experts, a context window of 1,048,576 tokens, and native visual support.

Independent third-party Artificial Analysis Intelligence Index v4.1 score: 57.1, ranked 4th globally.

The top three are Claude Fable 5 (59.9), GPT-5.6 Sol (58.9), and Claude Opus 4.8 (55.7), which is closely trailing Kimi K3. Below them, Grok 4.5 scores 53.8, GLM-5.2 scores 51.1, and DeepSeek V4 scores 44.3.

This is the first time a Chinese model has truly entered the global top tier.

Now let’s look at the specific sub-items: GPQA Diamond: 93.5%. Terminal-Bench 2.1: 88.3%, nearly matching GPT-5.6 Sol’s 88.8%. BrowseComp: 91.2%. On Arena.ai’s Frontend Code Arena, K3 took first place, outperforming Claude Fable 5. It also ranked first on AutomationBench-AA at 53%.

On the independent GDPval-AA v2 score, which measures "actual economic value knowledge work," K3's Elo rating jumped directly from 1,190 in the previous generation K2.6 to 1,668, surpassing Claude Opus 4.8 at 1,600, GLM-5.2 at 1,514, and GPT-5.5 at 1,494.

The jump was 478 Elo. Quite aggressive.

So where is the weakness?

High hallucination rate. Artificial Analysis found that K3 has a hallucination rate of 51%, higher than the previous generation. While accuracy has improved, the probability of making things up has also increased.

Too verbose. K3 generated 130 million tokens in the full AA benchmark, compared to the industry average of 63 million—more than double. Simon Willison tested a simple "draw a goose" SVG task, which cost about $0.25.

Open weights are delayed. On the release day, there was no checkpoint, no license, no model card, and no technical report. Moonshot has committed to releasing them by July 27—that is, six days from when this article was written. Until then, K3 can only be accessed via API and app, effectively making it a "hosted model."

No SWE-bench Pro/Verified score available. The evaluation sets selected by Moonshot, such as DeepSWE and FrontierSWE, cannot yet be replicated by third parties.

So, K3 is a model that is "truly good, but not yet perfect." This point needs to be clearly stated—we shouldn't exaggerate its strengths.

A more significant signal: China’s AI is no longer "as cheap as cabbage"

This is the most easily overlooked, yet potentially most important thing about K3.

K3 API pricing: $3 per million input tokens, $15 per million output tokens, $0.30 for cached inputs.

What was the previous generation Kimi K2.6? $0.95 / $4.00.

Increased by about three times.

What does this price mean? It’s essentially aligned with the Claude Sonnet tier. In other words, Moonshot has deliberately abandoned its most distinctive advantage over the past few years in China’s large model market—extremely low pricing.

Why?

Due to 2.8 trillion parameters, always-on deep thinking mode, and 1M context, the costs cannot be reduced. Additionally, K3 is no longer positioned as "an affordable assistant for you," but rather as an agent-level foundation model directly competing with Claude Opus and GPT-5.6.

Interestingly, even with a threefold price increase, Artificial Analysis calculates K3’s cost per task at approximately $0.94—about half that of Claude Opus 4.8 ($1.80) and nearly identical to GPT-5.6 Sol’s $1.04.

Half-price benchmark, same tier.

This is what caused the U.S. stock market to shake on July 17.

On that day, the Nasdaq fell more than 1.4%, and NVIDIA also suffered selling pressure. A product launch by a Chinese startup genuinely moved sentiment in the U.S. stock market.

This would have been unimaginable two years ago. At that time, the transmission pathway from Chinese model releases to U.S. equities barely existed.

It exists now because the narrative has changed.

The narrative has shifted from "America leads in cutting-edge development, while China follows" to "China also creates cutting-edge technology, and it does so with higher weights and at half the price." This new narrative directly, long-term, and structurally pressures the pricing power of closed-source frontier labs like Anthropic and OpenAI.

Behind the $20 billion valuation: A bet that quadrupled in three months

The financing timeline tells the same story.

By the end of 2024, Moonshot was valued at approximately $3.3 billion. By May 2026, a new round of funding totaling around $2 billion was completed, pushing the valuation above $20 billion.

Quadrupled in three months.

Yang Zhilin himself said at the Zhongguancun Forum in March 2026: "Over the coming years, an increasing number of research efforts will be led by AI." At GTC 2026, he further revealed three technical roadmaps: token efficiency, long context, and scalable expansion of agent clusters.

You need to understand these three points to grasp Moonshot's bet.

Token efficiency refers to the two new mechanisms in K3: Kimi Delta Attention and Attention Residuals, which reportedly increase decoding speed by up to 6.3 times.

Long context refers to the rapid progression from 200,000 characters to 1 million tokens.

Agent cluster, corresponding to K3, which is explicitly defined as an "agentic reasoning model"—designed for long-chain tasks, codebase navigation, and tool invocation, not for chatting.

These three paths represent Yang Zhilin’s bets on the next generation of AI.

He bets that a single large model is not the end goal—rather, it's a cluster of intelligent agents capable of dynamic generation and collaborative scheduling.

Whether this judgment is correct will be answered within two years.

Open science is China's true moat in AI.

A larger context must be mentioned here.

Recently, an observation from an overseas AI researcher has circulated widely within the community, roughly stating: Chinese labs are able to consistently produce models like GLM-5.2 and Kimi K3 not because they have more computing power, but because of openness—not just opening weights, but the entire ecosystem.

There’s a sobering observation here: Much of the model training work at China’s top labs is carried out by interns. Moreover, these undergraduate and graduate students have a deep understanding of training details and are 100 times more willing to share their knowledge than their American counterparts.

Conversely, leading U.S. labs rarely hire interns. Stanford and Berkeley PhD students must compete fiercely for a chance to secure computing resources to train models of "appropriate scale." Most of the secrets are locked away by a small group of privileged researchers.

This is not a contest between China and the United States. This is a contest between open science and closed science.

The very existence of Kimi K3 is the best illustration of this point.

A 34-year-old drummer and entrepreneur used his own paper as a foundation, leading a group of students or recent graduates to build the world’s fourth-largest model.

This is unlikely to happen within OpenAI—not because of the people, but because the system doesn't allow it.

Of course, K3's decision to delay the open-weight release until July 27 is itself a signal—the open-source camp is also hesitating, calculating, and being pushed back a step by commercial pressures. This move warrants caution, but it does not overturn the broader direction.

DeepSeek V4 remains open under the MIT license. GLM-5.2 is also open. The Chinese cohort of open-weight models has not disbanded.

How far can a drummer go?

Back to Yang Zhilin.

A tech entrepreneur with a background as a rock drummer, he’s almost an outlier in China’s AI circle. He doesn’t fit the typical mold of a "Tsinghua genius scientist" or the usual "Silicon Valley returnee elite."

He doesn’t make many public statements, but when he does, they’re dense with meaning. At GTC 2026, he laid out the technical roadmap with precision—restructuring the optimizer, reimagining the attention mechanism, rethinking residual connections. This isn’t PR jargon; it’s technical insight.

He also acknowledged Moonshot's shortcomings. K3’s own release blog was candid: compared to Claude Fable 5 and GPT-5.6 Sol, the UX still lags; it is sensitive to retained thought history; and it becomes "overly proactive" with ambiguous tasks.

An honest CEO, a researcher who turned their thesis into a product, a young team connected by Tsinghua University, and a clear technical roadmap—this is the foundation of Moonshot today.

It’s not perfect. It makes mistakes. Its path to commercialization isn’t fully proven—whether a threefold price increase can sustain growth, whether the computational cost of 1M context can be reduced, and when agent-based forms will truly materialize—these remain open questions.

But it's real. It's not a PowerPoint company, not a valuation bubble, not fueled by temporary hype.

Conclusion: The Nasdaq shake-up is just the beginning.

On July 17, 2026, the Nasdaq fell 1.4%.

Some might say, "What's a 1.4% drop? The market fluctuates every day."

But what you need to look at is the chain of causality: a Chinese startup releases a model, and on the same day, it is seriously reviewed by Reuters, CNBC, and independent evaluators like Simon Willison; it triggers a simultaneous sell-off of NVIDIA; and Artificial Analysis ranks it among the top 4 globally.

This has never happened before.

It means one thing: the pricing power and influence of AI are quietly shifting away from the San Francisco Bay Area—not entirely, but beginning to.

Yang Zhilin is not the only driving force—DeepSeek’s team is, Zhipu is, and Alibaba’s Tongyi is too. But he is undoubtedly the sharpest and most representative symbol of this round—a 34-year-old who understands theory, product, openness, and commercialization, and even knows how to play the drums.

There will be a next generation of the model. K4 and K5 will come eventually. Scores will always be tied and then surpassed.

What truly remains is whether this open path can be maintained.

If it's protected, AI won't be the AI of just three companies.

If we can't hold on to it, we'll simply move from one form of closure to another.

As for Yang Zhilin himself—he likely doesn’t care much about these grand narratives. At 34, with a company valued at $20 billion and a model ranked among the top four globally, surrounded by a team of young people he can truly connect with, he’s already in a position to keep drumming, keep writing papers, and keep being one of the few who turn their research into products.

When was the second time Nasdaq shook?

It probably won't take long.

Reference materials

1. Reuters, "China's Moonshot unveils the world's largest open AI model, closing in on US rivals", 2026-07-17
2. CNBC, "China's Moonshot AI unveils Kimi K3 that rivals OpenAI and Anthropic", 2026-07-17
3. AI Rankings: "Kimi K3: Benchmarks, Pricing & Review — Moonshot's 2.8T Frontier Model," July 2026
4. The Decoder, "Kimi's open model K3 nears GPT-5.6 Sol and Fable 5, signaling the end of super-cheap Chinese AI," July 2026
5. Artificial Analysis, "Kimi K3 Model Card & Intelligence Index v4.1", 2026-07
6. Wall Street View, "Kimi Founder Yang Zhilin: The Future of AI Development Will Enter an AI-Dominated Era, Company Valuation Reaches $18 Billion," March 25, 2026
7. The Age of Intelligent Minds: Yang Zhilin's GTC 2026 Talk: First Full Disclosure of the Kimi Model Technology Roadmap, 2026-03-18
8. China AI Atlas, "YANG Zhilin (Yang Zhilin) Profile"