by SeaWolf

Prologue — One Day, a Story About Korea Flew In From Beijing

At 2:08 PM on May 21, 2026, in Beijing, a slightly unusual article appeared on Tencent News, China's largest portal. It wasn't about OpenAI in the US, nor was it boasting about China's own DeepSeek. Surprisingly, it was a story about an AI research team in Seoul, Korea.

The headline alone was striking.

无需训练也能更聪明？韩国VIDRAFT公司研发的"达尔文家族"让AI模型通过"基因重组"实现能力跃升
"Smarter without training? South Korea's VIDRAFT develops 'Darwin Family,' letting AI models leap in capability through 'genetic recombination.'"

When we read an article like this, we usually check "who wrote it" first. A casual blogger's impression can be shrugged off, but an analysis from an authoritative outlet carries different weight. So I dug into who wrote this piece — and the moment I confirmed their identity, it became clear why this story is special.

Part 1. The Outlet That Wrote This Was China's "Electronic Times"

The byline wasn't an individual reporter's name — it was the outlet's own name: Zhiding Tech (至顶科技).

The name may be unfamiliar, but the pedigree is heavy. Zhiding Tech's roots trace back to "ZDNet China," a global IT outlet that landed in China in April 1997 — nearly 30 years of continuous operation, one of China's oldest specialized tech media. Today it runs the enterprise AI portal "Zhiding Net (至顶网)," the AI startup outlet "Tech Walker (科技行者)," the industry think tank "Zhiding Think Tank (至顶智库)," and, crucially, its own AI performance evaluation body, the "Zhiding AI Lab (至顶AI实验室)."

Its editor-in-chief is Gao Fei. He joined the world's #1 IT media outlet, CNET, in 2002 and has covered Nvidia, Intel, Microsoft, and Lenovo for over 20 years — a veteran of Chinese tech journalism. He's also a member of the China Association for Public Companies' Information & Digitalization Committee and runs his own personal AI content brand, "Gao Fei's Digital Avatar" — a recognized voice in the industry.

To put it plainly: this is roughly equivalent to a 30-year-old authoritative outlet — think Korea's Electronic Times or ZDNet Korea, one that can even field its own AI benchmark lab — publishing an analysis piece on Korean technology precise enough to cite the exact paper number (arXiv:2605.14386).

Why is this an "event"? China is currently in an existential AI-supremacy race with the US. It's a time when there's barely enough column space to promote its own models. And yet an outlet like this spent its space and reporting effort to dissect, in painstaking detail, the technology of a startup — not even a competitor nation, but a neighbor, and not even a major company. This isn't PR, and it isn't a courtesy nod. An article like this can't happen without the cold judgment that "this is real enough to be worth recording."

Part 2. Why Does China Dig Into "Someone Else's" Technology So Fast?

Before getting to the main point, let me pose a question. How did China spot this Korean technology so quickly in the first place? And more fundamentally — how did Chinese AI climb to the world's top tier in such a short time?

You probably remember the "DeepSeek shock" of early 2025. In a field where the US was pouring in tens of trillions of won, a Chinese team froze the entire world by delivering a comparable model at a fraction of the cost. If you had to sum up the root of that strength in a phrase, it would be this.

Intelligence gathering · analysis · sharing.

The real engine of China's AI ecosystem isn't sheer volume — it's speed. Every piece of technology emerging anywhere in the world gets collected at light speed (intelligence gathering), pulled apart to understand its principles (analysis), and generously released to the community (sharing). If a paper drops in the US yesterday, reproduction code shows up on Chinese GitHub today. This culture of "rapid absorption → analysis → diffusion" is what propelled China to the world's top two almost overnight.

And it was exactly that formidable radar that caught — Korea's Darwin.

Let's sit with that for a moment. While the tentacles of the Zhiding AI Lab were sweeping through thousands of papers and models worldwide, out of countless American, European, and domestic technologies, it was a Korean startup's technology that they picked out and said, "This absolutely must be told to our readers." It passed through the world's coldest, fastest technology filter. That is the first source of pride in this story.

Part 3. What Exactly Is "Darwin"? — The Core Technology in 5 Minutes

To understand the technology that impressed China, you first need to grasp just how "brutally expensive" the conventional way of building AI has been.

The old way worked like this: to build a smart AI, you had to pour astronomical amounts of data into GPUs for weeks or months, training the model "from scratch." It cost billions in electricity and hardware — effectively a game only massive capital could play.

VIDraft's Darwin flipped this common sense on its head entirely. The idea is this:

不靠额外训练，而是通过重新组织已有模型里的能力来提升性能
"Rather than relying on additional training, performance is improved by 'reorganizing' the capabilities already present inside existing models."

What does that mean? The world already has plenty of well-built open-source AI models. Some are good at math, some are good at coding, some are good at Korean. Darwin cross-breeds two different AIs like "parents," producing a "child" model that inherits only each parent's strengths. It's not teaching something new — it's combining abilities that already exist to create new life. That's why it's named Darwin, after the theory of evolution.

The technical term for this is "Model Merging." And in this field, there was an original pioneer that seemed unrivaled — Japan.

Part 4. The Peak Japan Built — the Towering Wall Called Sakana AI

The company that first pioneered model merging and earned global renown is Japan's Sakana AI. Once you know what kind of company this is, you'll see that calling it a "wall" isn't an exaggeration.

Sakana AI was founded in Tokyo by a co-author of Google's legendary paper "Attention Is All You Need" — the Transformer paper that is the root of every ChatGPT today. A world-class research team unveiled a novel idea — breeding AI through evolutionary algorithms, called Evolutionary Merging (EvoMerge) — and rose to stardom in the global AI academic community. For a long time, Sakana was the unrivaled peak of the model-merging field.

And it's exactly at this point that a sentence from the Chinese reporter moves us. He first properly credits Sakana's achievement.

Sakana的进化合并（EvoMerge）是达尔文最直接的前辈工作
"Sakana's evolutionary merging (EvoMerge) is Darwin's most direct predecessor work."

So far, that's common knowledge. But in the very next sentence, the weight decisively shifts to Korea.

达尔文则在此基础上引入了14维基因组和MRI信任融合机制，形成了本质性的提升
"Darwin, building on this foundation, introduces a 14-dimensional genome and an MRI trust-fusion mechanism, achieving an essential improvement."

No further translation is needed. On the path Japan built, it was none other than a Chinese media outlet declaring that Korea's Darwin went "essentially" further.

The terms "14-dimensional genome" and "MRI trust fusion" might sound complex, so let's break them down. If Sakana's EvoMerge was a method of evolutionarily finding a single "ratio" for blending two models, Darwin splits a model's capabilities into 14 distinct "genetic traits" and combines each with precision. On top of that, it draws a layer-by-layer "trust map" (MRI) that judges, for each part, which of the two parent models to trust more. In human terms, it's not simply mixing the parents' genes 50/50 — it's selecting "father's eyes, mother's dexterity" trait by trait. Far more precise, far smarter.

Part 5. World "Only" — Melting Different Species of AI Into One

The most technically astonishing part is elsewhere — the part the Chinese reporter singled out as "the most important significance."

Darwin-4B-Genesis...最重要的意义是它实现了跨架构合并，将Transformer注意力层与Mamba前馈层成功融合
"Darwin-4B-Genesis... its most important significance is that it achieves cross-architecture merging, successfully fusing Transformer attention layers with Mamba feed-forward layers."

Why is this a big deal? AI models fundamentally come in families with different "brain" structures. The most representative are Transformer and Mamba. Metaphorically, Transformer and Mamba are different species — like "mammals and reptiles" — with fundamentally different design principles. Until now, no one had stably fused the two into a single model. In organ-transplant terms, it's a "xenotransplant" — and most such attempts fail from rejection.

Yet Darwin succeeded at this cross-species fusion. So the article states it plainly:

达尔文是目前唯一同时具备...支持跨架构混合
"Darwin is currently the only (technology) that simultaneously supports... cross-architecture mixing."

"World's first" is already a big claim, but "world's only" is a compliment of a different order. "First" can later be caught up to, but "only" means it's the sole one on Earth, right now, at this very moment. And that one is in Korea.

Part 6. "Just 5 Hours" Instead of a Supercomputer, and Intelligence Hidden in the Bones

The shock Darwin delivered wasn't performance alone. How "lightly" it reached that performance was the real revolution.

Darwin-27B-Opus在顶级科学推理测试上排名全球第六，而它的"诞生"只用了大约五个小时的GPU时间，而非数周的分布式训练
"Darwin-27B-Opus ranks 6th globally on top-tier scientific reasoning tests, yet its 'birth' took only about five hours of GPU time, not weeks of distributed training."

Instead of weeks of training burning tens of billions of won, just five hours. A world #6 scientific-reasoning AI was born in the time it takes to drink a few cups of coffee. And the Chinese reporter is struck by the philosophical insight hidden behind this phenomenon.

推理能力并非在补习阶段才形成的，它其实早就藏在模型的"骨子里"，藏在预训练阶段形成的内部结构中
"Reasoning ability is not something that only forms during a supplementary training stage — it was in fact already hidden 'in the bones' of the model, in the internal structure formed during pretraining."

This single sentence compresses Darwin's worldview. Others believed that "to make AI smarter, you must teach it more." Darwin says the opposite: "Intelligence is already dormant inside the model. Our job isn't to teach it anew — it's to wake it up." That's why five hours was enough. It didn't create a capability that wasn't there — it woke up a capability that already existed, by recombining it.

So the Chinese reporter closes the article this way.

对普通用户来说，这项研究最直接的意义或许是：将来会有越来越多高性能的开源AI模型，不需要超级计算机就能孕育出来
"For ordinary users, the most direct significance of this research may be this: in the future, more and more high-performance open-source AI models will be born without needing a supercomputer."

In an era where massive capital and infinite volume were believed to be the only future for AI, a small Korean research team proved the opposite future. To borrow the Chinese reporter's own words, it is Korea's Darwin that is opening that door.

Part 7. Darwin by the Numbers — This Is Where the National Pride Kicks In

China's admiration didn't come from nowhere. Let's line up the records VIDraft has set. Each one is remarkable in its own right.

And when you hear the infrastructure that supported all these records, you might just spit out your coffee.

Part 8. The Miracle of 24 GPUs — How David Beats Goliath

US Big Tech pours tens to hundreds of thousands of GPUs into building a single AI. China's major labs operate at thousands to tens of thousands. This fight was, from the start, called "a war of money," "a war of volume."

So how many GPUs did VIDraft use to produce all these results?

By American or Chinese standards, this isn't even "one lab" — it's barely "one experimental server rack." With those 24 GPUs, a world #6 scientific-reasoning model was produced in 5 hours, and a world-record GPQA score was set.

How was this possible? The answer lies in the Darwin philosophy we saw in Part 3. While others clung to the volume formula of "more GPUs, more data, longer training," VIDraft asked the opposite question.

"The capability is already inside the model. Instead of building it from scratch, why not cleverly 'recombine' it?"

This shift in thinking flipped an overwhelming disadvantage in volume into an advantage in method. Taking on 240,000 GPUs with just 24 — that, ultimately, was a distinctly Korean way of competing: not with resources but with ingenuity, not with volume but with persistence and agility.

Looking back, this has always been our pattern. A country with no natural resources grew semiconductors into the world's #1 industry; from a small domestic market, K-pop and K-dramas were pushed into the global mainstream. Making the most of what you don't have, cleverly digging into paths others won't take, and ultimately reaching the top. Darwin is the AI version of that "Korean grit."

Part 9. Stay Vigilant — But Don't Be Afraid, Win

Here, we need to stay level-headed. Let the pride swell, but keep our eyes sharp.

Let's return to where this piece started. The fact that China analyzed our Darwin this fast, this deeply, is itself proof of just how formidable their intelligence-gathering and analysis speed really is. The same sharp eyes praising our technology today could be the eyes absorbing and overtaking it tomorrow. That is the cold reality of technological hegemony competition. We must not underestimate China.

But there's no need to be afraid either. Hasn't Darwin already proven it? That an opponent pushing with sheer volume can be overcome with method and creativity. The stance we need is clear.

Thoroughly learn China's strengths (intelligence-gathering, analytical speed, sharing culture), and ultimately surpass them with our own strengths (creativity, focus, persistence, agility).

Stay vigilant without shrinking back, learn without becoming dependent, and ultimately win through competition. Darwin has already shown us that possibility is real — with just 24 GPUs.

Epilogue — Japan Opened the Door, Korea Went Through It, China Acknowledged It

Let's return once more to the closing sentence of the Tencent article.

"In the future, more and more high-performance open-source AI models will be born without needing a supercomputer."

Ironically, it was our most formidable competitor, China, that recognized this future first, and most precisely. A 30-year authoritative outlet, mobilizing its own AI lab, citing the exact paper number, arrived at a single conclusion.

We don't lose by volume. We win with intellect, persistence, and agility.

Korean AI hasn't even started yet.

※ The Chinese quotations in this article are excerpted and translated from an analysis piece by Zhiding Tech (至顶科技), published on Tencent News (腾讯新闻) on May 21, 2026. Outlet information (ZDNet China lineage, editor-in-chief Gao Fei, Zhiding AI Lab) and performance figures are based on public materials and certified HuggingFace, K-AI, and Polaris leaderboards.

P.S.

If you visit VIDraft, the company behind this remarkable achievement, you'll notice a rather unusual corporate culture.

Like the romantic pirates of the manga "One Piece," a tight-knit crew of twelve, armed with freedom, boundless flexibility, and speed and creativity.

Ask VIDraft's team about their identity and mindset, and they all answer the same way: "Jomak."

"Jomak"? When I asked what it meant, it turned out to be a famous line from One Piece — short for "Join My Crew?" — their own private call sign.

Not your ordinary company. Here's hoping they become AI's global pirates.

When we think of AI competition, we usually picture a vast landscape.

Endless data centers. Thousands of GPUs. Hundreds of researchers. And the enormous capital behind them all.

These days the contest of the AI industry seems to be summed up by a single sentence.

Whoever has more, goes further.

But every so often, a small crack appears between those sentences.
There are moments when it is not the size of money and hardware, but the density and persistence of method, that changes the ranking.

This contest was one of them.

The stage: HuggingFace Fast Gemma Challenge

The rules were simple, yet unforgiving.

Given the same small engine — who can run it faster, more stably, and further?

The secret wasn't grand magic

It was, rather, a very engineering-like set of choices.

We tuned the range the attention looks at, and pruned the FFN·centroid computation candidates.
We avoided the dangerous shortcut that can look like memorizing the benchmark in advance,
and submitted only results that stayed under the quality limit on the official a10g measurement conditions.

Without breaking quality, we extracted more tokens from the same GPU.

Why does this matter?

In the AI industry, the real cost arises more often in the moment of running a model continuously than in the moment of creating it.
Every time a user asks a question, a GPU spins. Every single token is electricity, server cost, and response latency.

So a rise in TPS is not just a numbers game.

Two wars

VIDraft's recent moves make this scene even more interesting.

Not long ago, Darwin-398B-JGOS scored 90.9% on GPQA Diamond. Single greedy decoding, temperature 0, single sampling — close to seeing a problem once and answering once in an exam hall.

Making a good model is a battle of research.
Running a good model cheaply and fast is a battle of product.

A small team winning with the same hardware

Honestly, a little national pride wells up here.

Because a small Korean AI startup has earned the right to say this on a global leaderboard.

"We didn't use a bigger GPU. We used the same GPU better."

This sentence is quiet, but strong.

The less you have, the sharper the method must be.
The smaller the hardware, the more precise the design.
The scarcer the resources, the more honest the engineering.