January 20, 2025. No launch event.

No countdown. No founder walking on stage. No word like "redefine." A repository gained a weight file and a technical report — several dozen pages, written like a paper, not like an ad.

Six hours later, it was the top free app in the U.S. App Store.

Seven days later, on January 27, a company that sells shovels fell 16.97% in a single session, shedding $592.7 billion in market value. The largest single-day loss any American company has ever recorded.

It wasn't defeated. It was simply shown that maybe you don't need that many shovels.

To explain this, you first have to explain a place.

Not a company. Not a lab. A library on the back hill of Shaolin — one that builds no gate, takes no disciples, enters no tournaments. It only sweeps, copies, and carves its methods into the wall.

And the man who keeps it never attends the banquets.

I. The Man Who Sweeps

1985, Wuchuan, Zhanjiang, Guangdong. A village called Mililing.

Both parents taught Chinese in a primary school. In 1990s Guangdong, money was everywhere and books were not. Parents used to visit his father to discuss their children's prospects, and the conversation always landed in the same place: studying is useless.

That was the air he grew up in.

In the first year of middle school, he finished the high school math curriculum. Then he started on university math.

Notice what kind of act this is: nobody asked him to. No teacher assigned it. No competition syllabus required it. No exam bonus rewarded it. A kid from a county town in western Guangdong worked two or three grade levels ahead on his own — that isn't doing the exercises. That's finding the problems when nobody has set any.

2002: top scorer at Wuchuan No.1 High School, admitted to Zhejiang University for electronic information engineering.

2005: first prize in the National Undergraduate Electronic Design Contest. Worth isolating: that is skill in the hands, not points on paper.

2008, global financial crisis. Still a graduate student at Zhejiang, he and classmates began accumulating market data and experimenting with machine learning for fully automated quantitative trading. Initial capital: 80,000 yuan.

That year, AlexNet was still four years away. "Artificial intelligence" lived in papers and science fiction. He told people AI would change the world. Nobody believed him.

His master's thesis: Research on Target Tracking Algorithms Based on Low-Cost PTZ Cameras.

"Low-cost" — those two words are the master plan of his entire career. Twenty years later, they would become $5.576 million.

A reporter once asked why he hires for ability rather than experience. His answer is essentially a self-portrait:

"Experienced people tell you without hesitation: this is how it's done. People without experience grope around, think carefully about how it should be done, and then find a solution that fits the actual situation."

Nobody opened his mind for him. He opened it himself.

II. The Field Behind the Library

2015: High-Flyer is founded.

The name comes from the Luoshu magic square — an arrangement of numbers where every row, every column, and both diagonals sum to the same value. From the day he named it, what he wanted was balance, not maximum.

October 2016: High-Flyer produces its first trading position generated by deep learning. By the end of 2017, nearly every strategy runs on AI.

Then comes the thing his peers could not understand at all — buying GPUs.

2019: nearly 200 million yuan for "Fire-Flyer I," 1,100 GPUs. 2021: one billion yuan for "Fire-Flyer II," roughly ten thousand A100s.

The field grows no grain. It grows compute.

Why would a quantitative hedge fund hoard that much compute?

The jianghu guessed at hidden business logic, at an arms race. His own explanation:

"People assume there's some secret commercial logic here. But it was mainly curiosity."

Then he offered an analogy — the best one I have read in any interview with a Chinese technologist:

"Something thrilling can't always be measured in money. It's like buying a piano for the house — first, you can afford it; second, because there are people eager to play it."

A year later, export controls landed.

By 2023, apart from a handful of internet giants, the only outfit in China holding ten thousand A100s was High-Flyer.

What looked like a quirk had become one of Chinese AI's earliest reserves.

Here the story wants to become a triumphant tale: look how far-sighted he was.

Let me flip it.

This was not foresight. It was the curiosity of someone who could afford it. The same act, done by someone without money, is called gambling.

So the conclusion of this chapter isn't "he saw it coming." It's this:

High-Flyer is that field. It doesn't grow grain. It grows "ten years without having to explain yourself to anyone." Every unreasonable thing DeepSeek later did has its root in that field.

III. Sheathing the Blade

May 2024. DeepSeek-V2 is open-sourced.

It did what nobody had done: pushed input pricing to one yuan per million tokens — one percent of GPT-4's.

Five days later, Zhipu matched it. Then ByteDance, Alibaba, Baidu, Tencent. China's price sheet was rewritten.

And note: only China's. The shot was fired indoors. Beyond the wall the price sheet did not move an inch — OpenAI's rates held, and nobody in Silicon Valley turned around. What made the other side of the Pacific begin to sit up arrived seven months later.

The technique underneath is called MLA — Multi-head Latent Attention.

Plainly: when a model talks to you, the expensive part isn't striking. It's remembering every move you've made. That memory is the KV cache. The longer the conversation, the more it grows — linearly with length. Long context is expensive almost entirely because of this.

MLA's move: compress the trace of each move into a single faint mark, sheathe it, and unpack it only when needed.

Thirty-six weapons' worth of memory, in a quarter of the space.

The blade stays sheathed; the edge lives inside.

Call it the Art of the Hidden Edge — the blade stays sheathed; the edge lives inside the sheath.

People called him a catfish stirring up the pond. His denial was better than the accusation:

"We didn't set out to be a catfish. We just accidentally became one."

Asked about pricing logic, he was plainer still:

"Our principle is no subsidies and no windfall profits. The price carries a small margin above cost."

One fair point in his defense.

That cut hurt him too. The giants could afford it — cloud businesses, ad revenue, ecosystems; they could treat a model as a loss leader. He had none of that. He had a field.

So he had to actually bring the cost down. Not a slogan. Not burning cash for share.

This wasn't generosity. It was a technique forced out of him.

IV. Splitting the Light

If V2 was about hiding, V3 was about dividing.

December 26, 2024. V3 is open-sourced, and the technical report carries a line that made Silicon Valley sit up:

To see how sharp that is, put it beside the others:

Sect	Weapon	Training cost	Iron used
The Library	V3	$5.576M	H800 (export-restricted)
Shaolin	GPT-4o	~$100M	H100
The other open sect	Llama 3.1 405B	>$60M	16,000 GPUs

One blade against a mountain of scrap iron. The scale is level.

It didn't sharpen the blade. It forged the same blade with a tenth of the iron.

Three techniques.

One: splitting the light

The point isn't saving. It's dividing. A trillion units of qi split into hundreds of avatars; when a fight comes, only a few dozen wake up. The rest sleep.

The clever part is the "shared expert": a set of avatars permanently assigned to common knowledge, so no avatar has to relearn that one plus one is two.

The original body stays still; the avatars meet the enemy. Total parameters can reach 1.6 trillion, while only 49 billion work per token.

Two: three lives ahead

Predict several future tokens at once during training, and one dataset yields three times the signal.

The side benefit: at inference it doubles as speculative decoding, so it emits faster.

Three: the formless fire

Train at lower precision, save half the memory and bandwidth. This is not something you can simply decide to do — drop precision carelessly and training collapses. He lit that fire only after extensive numerical validation.

But the real kill shot was the price sheet

The technical report is about what it cost to train. What made people sit up was what it cost to buy.

In the December 26, 2024 announcement, DeepSeek wrote a line that read almost like a boast: "Until Feb 8: same as V2!"

After February 8: $0.27 per million input tokens (cache hit $0.07), $1.10 for output.

At that same moment, GPT-4o stood at $2.50 and $10.00.

Nine times cheaper on input, nine times cheaper on output — 89% cheaper overall. And its benchmark scores sat within reach of GPT-4o.

That is no longer "cheap." It was the first time anyone tied equal capability and one-tenth the price onto the same sheet.

So the R1 strike three months later didn't come out of nowhere. It stood on a cost structure V3 had already proven.

Now, something less flattering.

People have never quite believed the $5.576 million. Their argument: that's probably just the electricity for the final run.

That objection is correct.

It excludes dozens of failed experiments. It excludes the opportunity cost of ten thousand GPUs. It excludes people — a team averaging under 26 years old, whose cost is nearly unpriceable.

And even so, the gap remains an order of magnitude.

V. The Shout

January 20, 2025. R1 is open-sourced.

Its significance isn't performance. It's the discovery that something else is possible.

The traditional method: show the student thousands of worked solutions with standard answers, let them imitate. That's supervised fine-tuning. The problem is you can't tell when the student has learned the wrong thing — they learn to please the teacher, not to solve the problem.

DeepSeek did something severe: it stopped teaching.

R1-Zero skips supervised fine-tuning entirely. Only the final answer is scored. Right, reward. Wrong, penalty. How it thinks in between is not its teacher's business.

Then, at some intermediate checkpoint, the model stopped mid-solution and wrote:

"Wait, wait. Wait. That's an aha moment I can flag here."

Then it went back and walked through the entire solution again.

The pen stopped in mid-air. The report calls it an aha moment.

DeepSeek's technical report calls it the "aha moment." The wording is restrained, but the meaning is clear: we didn't teach it how to solve. We gave it the right incentive, and it grew a better method by itself.

The memory trick is just as clever. It's called GRPO: no referee model. A group of answers to the same problem are compared against each other and rewarded by rank. Remove a referee the size of the contestant, and half the arena empties out.

What was shaken that night was not a leaderboard.

It was the doctrine that more compute necessarily means more capability.

For ten years the whole jianghu believed one law: ten times the GPUs, ten times the power. The Library slid over a note with three words on it: not necessarily.

On January 27, the app topped the U.S. free chart. The same day, the shovel-sellers recorded their largest single-day loss ever.

What actually forced Silicon Valley to rewrite its price sheet overnight was the comparison: R1's input pricing ran at under one-thirtieth of o1's. For the same million tokens: about $15 on one side, $0.55 on the other.

For the first time, the global price anchor of AI was not set in Silicon Valley. OpenAI's later free release of o3-mini, its cuts to o1, and everything that followed trace back along this chain — but the source is here.

Marc Andreessen wrote on X the line that has been quoted ever since: "DeepSeek R1 is AI's Sputnik moment."

VI. Carving Sutras into Stone

The first five chapters are about technique. This one is about how technique is transmitted — and this is the hinge of the whole piece.

September 2025. The R1 paper appears on the cover of Nature. Liang Wenfeng is corresponding author.

Weights, methods, training details — all laid on the table. In the years when frontier labs were tightening every detail, a Chinese lab handed its recipe to the world for inspection.

Carved into the wall: no master to bow to, no tribute, no attribution.

Why?

First, the unromantic answer

Rewind to December 2024. A Hangzhou company, a couple hundred people, claims it built something close to the Silicon Valley frontier in fifty-five days, on two thousand cards, for a little over five million dollars.

What happens if it stays closed?

No one can verify it. No weights, no training logs, no third party who can rerun the thing. Every judgment reduces to believe or don't — and in this industry, belief demands collateral. Microsoft trusts OpenAI because it invested over ten billion dollars. Enterprises trust Anthropic because AWS and Google stand behind it. What does an unknown Chinese company put up as collateral?

The doubt never dies; it only ferments. First "they copied it," then "they faked it," then "they can't even show the thing." None of these three need evidence — only time. And time is exactly what he lacks. Models turn over every few months; by the time he finishes answering each accusation, the table has changed hands. By then, falling behind reads as confession: "See — it was true all along."

The harder wall sits on the procurement table. For a big overseas enterprise to adopt a foreign model's API, what must it pass first? Where the data flows, who answers when something breaks, whether a local entity can be held liable, whose signature goes on the compliance report. These are not gates a benchmark score can pass. They are built from company size, legal teams, regulatory relationships, years of track record. He has none of it.

And open source is the only door that walks around all of it.

Because once the weights can be downloaded, the cloud providers can move them into their own data centers: data stays in-country, liability sits with the cloud, the compliance chain stays in their own hands. In January 2025, Microsoft put R1 on Azure, AWS put it on Bedrock, Perplexity self-hosted it on U.S. servers. It was the cheapest way "a Chinese startup's product" has ever entered American mainstream infrastructure.

So open source wasn't him being accepted — it was him going around. Around that pile of trust assets he could never accumulate. Not by proving himself trustworthy, but by making it so no one has to trust him in order to use him.

There is an almost clean control group. The most successful Chinese models abroad are all open-weight: DeepSeek, Qwen, GLM, Kimi K2, MiniMax. The closed-source Doubao wins at home on distribution, yet barely exists among overseas developers. The most telling is Baidu — ERNIE was closed source as a matter of identity, and in June 2025 Baidu open-sourced ERNIE 4.5 whole. Even the company that least needed to, changed course.

But that's only one side of the ledger. He gave another answer himself, and it doesn't sound like a businessman's.

In that 2024 interview, the reporter asked directly: where's your moat? MLA will be copied soon, won't it?

His answer is, I think, the most quoted and most under-read passage in Chinese AI:

"In the face of disruptive technology, a moat built on closed source is brief. Even OpenAI, being closed, cannot stop others from overtaking it."

"So we deposit value in the team... an organization and culture that can innovate — that is our moat."

"Open-sourcing, publishing papers — we lose nothing. For a technical person, being followed is deeply satisfying. Open source is more a cultural act than a commercial one."

Asked whether they would ever close it: "We will not close-source."

"A cultural act" is true. It is also incomplete.

Let me supply the half he can't say out loud.

Open source is the smartest move available to the weaker party.

You have less capital, less distribution, less ecosystem. So don't fight on the battlefield they chose. Move it to trust and default.

• Free weights destroy the price anchor. Anyone charging ten dollars per million tokens now has to justify it.
• But "the whole world writes code, calls tools, and fine-tunes on your architecture" is a position money cannot buy.
• The MIT license means: no bowing to a master, no tribute, no attribution. Copy it, take it, modify it, and go found your own sect.

He isn't giving away martial arts. He is changing where the world learns martial arts.

And here is the deepest layer: a sect's worst fear is not being copied. It is being forgotten.

The more it is copied, the heavier the mountain gate. Because once people are used to these methods, the next edition of the arms catalog has to be dimensioned to them. By then, "who is the orthodox lineage" is a meaningless question — everyone is built on your bones.

But above this layer sits another, running the other way. It is less flattering — and it is where the real weight of the whole thing lies.

And the price is far heavier than "clever"

Carving the sutras into stone is not a move. It is a long-term debt.

A closed moat is "they don't know" — and as long as the secret holds, it can hold for a long time. An open moat is "they can't catch up" — and that moat has to be re-earned every single day.

Because the moment the weights are out, they become the industry's starting line: others fine-tune on them, distill from them, reshape their own architecture off yours. And an advantage spreads faster than it is created. Training a model takes months; learning its weights takes days.

So the open-sourcer carries a constraint the closed camp never has: you must keep running ahead. Otherwise what you give away becomes the ammunition used against you.

That takes not cleverness but nerve — a very particular kind of nerve. Fail closed, and you can quietly retry, no one knowing you ever retreated. Fail open, and the whole world is training on your last release; your falling behind is public, itemized, used as a stepping stone. Every release is a risk of public execution.

And he walked straight into exactly that constraint.

After R1, R2 didn't ship for eighteen months. In those months, "reasoning can grow from reinforcement learning alone" was learned by the entire world, written into everyone's training pipelines, and turned into an industry default. The method he once carved into the wall became everyone's foundation. Then the third-party assessment came out: V4 is roughly eight months behind the American frontier.

By the "ammunition" logic, that should have been the end of him.

It wasn't. In 2026, DeepSeek still held about a quarter of OpenRouter requests, and on Ramp's U.S. enterprise spend index it was still the top software vendor. Eight months behind, and the seat was still his.

That is the insurance open source gives back to him: once your weights become someone else's infrastructure, falling behind stops meaning replacement — because replacing costs more than catching up.

So the real constraint is not "always number one" — no one can do that. It is "never drop out of the first tier." The first is heroism; the second is a discipline you can actually execute.

He kept the second, lost the first. And that is what "bravest" really weighs: not being unafraid — but laying the cards down knowing he would lose a stretch of the race.

So carving the sutras was never a safe move, and never a kind one. It was the most dangerous move — and the one most likely to make a name. Dangerous, because every release must pass in public and every stumble is a face-slap; likely to make a name, because an unknown can carve himself into the entire jianghu's foundation without waiting for anyone's permission.

And once that step was taken, the library's seat could never be moved again.

VII. Three Ways of Seeing

April 24, 2026. V4 ships.

Its best idea comes from an old line about Chinese landscape painting:

Take in the far distance as shape, the near field as form, and study closely only the few far things that matter.

A model handling long text fears one thing above all: needing to see everything clearly. A million tokens held at uniform precision will blow out memory.

V4's answer is to hold the same passage at three precisions at once:

• Distant shape: compress a million tokens into roughly eight thousand coarse entries. Knowing "what this is broadly about" is enough.
• Near form: loosen compression to four-to-one, then pick only the few hundred entries most worth reading and study those closely.
• Immediate field: the last hundred-odd tokens, held unchanged.

Far as shape, near as form, close only on what matters.

Plus two new pieces:

• The stabilizing seal: constrain residual mappings to the manifold of doubly stochastic matrices, so signal can't run wild — a guard against going berserk in deep training. The spectral norm is pinned at 1.
• The bone change: replace AdamW with Muon for faster, steadier convergence.

The numbers are hard. V4-Pro has 1.6 trillion total parameters and activates 49 billion; at a million-token context it needs 27% of the inference compute and 10% of the KV cache of the previous generation. Against a conventional GQA-8 baseline, the cache is about 2%.

This release isn't about saving money.

It's about turning long context from a luxury into a utility.

And why now? Because agents changed the workload. An agent doing one job reads dozens of files and runs dozens of tool calls, over and over. Long context used to be for demos. Now it's for work.

But I won't only praise.

V4 also shows the strain.

The technical report lists Huawei's Ascend NPU alongside Nvidia GPUs, with FP4 specifically tuned for Ascend 950. That's a step toward digging a new well — moving the water source out of someone else's yard.

The well isn't through yet.

R2 still hasn't shipped — eighteen months after R1, long enough for the jianghu to half forget the name. According to reports, DeepSeek was encouraged in 2025 to train R2 on Ascend; months of failed runs followed — instability, insufficient inter-chip bandwidth, an immature software stack — and it reverted to Nvidia for training while keeping Ascend for inference.

Meanwhile, third-party evaluation puts V4 roughly eight months behind the U.S. frontier. Domestic monthly actives sit at 130 million, behind Doubao's 382 million and Qwen's 167 million.

VIII. The Box of Ten Thousand Parts

August 13, 2026. DeepSeek open-sources again.

Not a model this time. The box the model goes in.

Harness. Developer preview v0.1, MIT license. Fifty thousand GitHub stars in twelve hours — one of the fastest-growing projects ever.

Its purpose hides in DeepSeek's own recruiting line, a formula of four words:

Not an all-purpose sword. A box anything can plug into.

The model is the brain — it thinks. Harness is the hands and the nerves — it schedules tools, tracks task state, closes the execution loop, and turns thinking into finished work.

Its most distinctive design principle: everything is a plugin. Built on Cordis, model, tools, skills, sessions, sandbox, storage, loops, scheduling, and UI are all plugins — all replaceable.

You never touch the box, and you can swap out anything inside it.

One more feature I think is undervalued: the session log is append-only. Everything the model saw — system prompt, chain of thought, tool calls and results, sub-agent dispatch — is recorded. In the trajectory view you can trace by source, resume, fork, search, replay.

In plain terms: it gives the agent a flight recorder. When it fails, you can find the step where it went wrong instead of guessing at a black box.

Now the real weight of this chapter.

One assessment gets to the bone: open weights answered "who can run the model"; open Harness reaches "who decides how the model works."

The first open source handed out the weapon.

The second handed out the manual.

Because giving a sword is a concession. Giving the manual is surrendering a paradigm. Once the world's agents grow on "everything is a plugin," what he defines stops being a model and becomes a way of working.

IX. No Patron Gets a Vote

The first eight chapters were about technology and philosophy. This one is about money.

Because without it, the first eight are fairy tales.

In 2026, DeepSeek changed.

A first round of roughly 51 billion yuan, of which Liang personally contributed about 20 billion. A second round reportedly approaching 100 billion, pushing valuation toward 500 billion yuan. Preparation for a STAR Market listing. Annualized revenue in the range of a billion dollars.

Two years earlier he had said the company wasn't in a hurry to raise, that money was never the problem.

He broke his own early vow.

But this isn't hypocrisy. It's an admission of constraint: idealism has a physical ceiling. Paper costs money. A GPU does not get cheaper because you care.

What matters is how he took the money.

Most outside capital went into limited partnerships where Liang is general partner — financial return rights only, no voting rights, shares locked for five years.

He holds roughly 84% and close to 100% of voting power.

Money waits at the door. The key hangs at the sweeper's waist.

In jianghu terms:

A patron may share in the incense money. A patron does not decide how the sutra is read.

Almost no founder designs this structure. Because most founders don't need money — they need the resources and endorsement behind it. And resources and endorsement usually arrive carrying "you'll listen to me."

He didn't want that. He wanted only the money.

Why does this matter so much?

Because a technical person's deepest fear was never poverty.

It's being told, once there is money, "this quarter, you need to monetize."

The quarterly ledger is the most reliable binding spell in the world. More diligent than any ideal, because it is recited every three months. Countless technical visions have died on that sentence.

So let me say something unfashionable.

The real dividing line is not technical person versus business person.

It's whether you own a field.

Liang could do all those uneconomical things because High-Flyer was feeding him — no ratchets, no exit timeline, no board to explain why this quarter had no revenue.

That is DeepSeek's real singularity: not "a technical person building models," but "a technical person who doesn't need to raise money, building models."

The former is everywhere. The latter is almost only him.

X. The Keeper Never Descends

The sweeping monk of Shaolin's library is the strongest figure in Jin Yong's fiction.

He never leaves the mountain. Never gives lectures. Never contests for abbot. He sweeps and turns pages. And when he finally speaks, the abbot and two senior monks concede on the spot.

Because all the martial arts in the world had always flowed out of that room.

Liang resembles him — but only half.

First, he is not a test-taker

"Test-taker" carries a clear implication: shaped by exams, good at solving problems others set, unable to pose one himself.

His record runs the other way.

Teaching himself university math in middle school — nobody set that problem.
Deciding in 2008 that AI would change the world, when nobody believed it — that problem did not yet exist.
Winning the electronic design contest in 2005 — that's a problem for the hands, not the page.
Building a model in 2023 because "this is a question of how, not why" — he rewrote even the question.

Second, he is not a god

R2 is eighteen months late. Eight months behind the frontier. Third in domestic monthly actives, behind Doubao by nearly three times. Raising a hundred billion, preparing to list.

His "grand plan" is not prophetic design.

It is finding the optimal move inside the constraint, again and again:

• Denied compute — turn efficiency into a moat.
• Outgunned on distribution — turn open source into the default.
• Needing money — lock the money outside the decision room.

That is harder than omniscience, and far more credible. The former needs intelligence. The latter needs fifteen years of not drifting.

Third, he is the rarest kind of keeper

He carved the most valuable thing onto a wall and let anyone copy it for free.

Then he used the fact that the more it is copied, the heavier the mountain gate becomes — to occupy a position no one else in the jianghu can occupy.

In Closing

Everything above, in seven lines:

1. It is not a sect. It is terrain. Others compete on whose model is stronger. It competes on where the world gets models. The first is a ranking. The second is geography.
2. Open source is not charity. It is the sharpest weapon of the weaker party. Unable to beat your capital and distribution, it takes your pricing power and your standard instead.
3. $5.576 million is not a stunt. It is a verdict. Not "who is cheaper," but that more-compute-equals-more-capability has expired. That verdict outweighs any leaderboard.
4. The field matters more than the technique. Without it, Liang is just a clever founder. With it, he can go ten years without explaining himself.
5. The second open source is the real move. Weights let people use it. Harness makes people work the way it works. The first concedes margin. The second sets the paradigm.
6. But the iron isn't his. R2 delayed, eight months behind, Ascend training that failed — these aren't blemishes, they're the ceiling. Every cost cut to the bone except the one that can't be.
7. He is not a test-taker. He is a problem-setter. And that applies beyond one man: for thirty years, Chinese technologists solved problems others set. Now some have started setting them. Only the problem-setter gets to change the rules.