China's Strategy to Win the AI Race

0:00 On a Sunday last week, a folder appeared on HuggingFace with no announcement attached to it at all. 133 files, 577 gigabytes of weights just sitting there. 280 billion parameters published under

0:14 Apache 2.0. Take it, fork it, sell it. 5 days later on the Friday, the account that uploaded it finally posted. The account is called Dots Studio. Dot Studio is the AI department of Rednote

0:27 at home. Shiao Hungshu. The app 400 million people open every day. So, China's Instagram shipped a frontier class agent model, gave it away, and said nothing about it for 5 days. 16

0:39 billion of those 280 billion parameters switch on for any given token. The rest stay asleep. Half a million tokens of context and four kinds of input. Text, images, video, and audio, all going in

0:52 at once. And is built for agents that keep going. their own test harness allows 10 hours per task before giving up. The reflex read was easy enough. Another Chinese model, another week, add

1:03 it to the pile and move on. But a photo sharing app does not spend Frontier money by accident. So what does this one know that the rest of us have not noticed yet because it is not one app.

1:14 On the 28th of April, Xiaomi, the phone company, released Mimo version 2.5, 310 billion parameters MIT license waits on the table. On the 30th of June, Mtoan, which delivers food, put out long cap

1:29 version two, 1.6 trillion parameters, also MIT, trained on Chinese chips, a food delivery company, handing out a trillion parameter model. Tencent followed on the 6th of July with HY3.

1:42 Moonshot on the 16th with Kimmy K3, 2.8 trillion parameters, Ant Group on the 27th, then Red Note on the 14th of August. phones, food, games, payments, photos, five unrelated businesses in

1:56 under four months. The one obvious name missing is Bite Dance, whose seed models stayed behind an API that absence tells you the giveaway is a choice. Watch what the choice did downstream. Open router

2:07 is the neutral router many developers buy tokens through, and it publishes the split. In June of last year, the American labs had about 70% of its traffic. 12 months later, they had 30.

2:17 Chinese openweight models now run about 61% of the tokens on that platform. Alibaba's Quen family passed a billion cumulative downloads and sits underneath 113,000 derivatives, forks of forks of

2:29 forks. The talent map explains more than people expect. 38% of the world's top AI researchers did their undergraduate degree in China. At the leading American institutions, 38% of the talent is

2:42 Chinese educated against 37% American. Among Silicon Valley startups running open source stacks, about 80% build on a Chinese-based model. So, is Beijing paying for all of this? Start by

2:54 admitting the state money is real, specific, and large. In January of last year, China stood up a national AI industry investment fund worth $8.2 billion, plus a venture guidance fund

3:05 with a headline size of a trillion yuan. Some inland provinces cut data center power bills by half. And last August, the state council published the AI plus action plan, which sets adoption targets

3:16 the way earlier plans set steel targets. 70% penetration of AI agents and smart terminals by 2027 over 90% by 2030. But none of that is a check with this model's name on it. No program I could

3:29 find attaches to DOT 3, and you do not need one because Rednot's own accounts cover it comfortably. Revenue last year came in near 42 billion yuan, up about 40%. And for the first time, most of it,

3:42 roughly 76% was advertising rather than commerce. Net income cleared $2 billion. That is what paid for the lab. Before this year, it was about a h 100red people. And its first open model back in

3:54 June of last year was half the size of this one. On the 30th of April, the company promoted it from a research group into a first level department reporting straight to the president. The

4:03 reason its own engineers give has nothing to do with policy. It is one sentence repeated internally. Without this capability, you cannot get a seat at the table. Fear predicts corporate

4:14 behavior better than a subsidy does. But fear explains the spending, not the giving away. Why hand the finished thing to your own competitors? For that, you need a software essay from 2002. Joel

4:25 Spolski, then running a small software company, wrote a piece called strategy letter 5, and its whole argument fits in one line. Demand for your product goes up when the price of the thing next to

4:36 it goes down. Cheap flights fill hotels, so smart companies work to make their compliments worthless. IBM pushing open source to sell consulting. Netscape giving away the browser to sell the

4:46 servers. Red Note sells attention and commerce to 400 million people. Intelligence is the compliment. If intelligence stays expensive and owned by three American labs, Rednote pays

4:56 rent forever on the feed it already owns. If intelligence is free, it keeps the feed, keeps the commerce, and keeps the margin. Giving away a model in that position is not generosity. It is the

5:08 cheapest insurance a company with $2 billion of profit will ever buy. And China has run this play in physical goods for 35 years, which is why it looks familiar from outside. In 1990,

5:20 Chinese mills made 8.6% of the world's steel. Last year, they made 54 12. Solar modules cost 60 cents a watt in 2010. Top tier Chinese panels traded near 9 cents this year, an 85% collapse. While

5:34 China took over 85% of world module production, and its own makers posted a billion and a half in quarterly losses. Lithium battery packs cost $1,474 a kilowatt hour in 2010 in today's

5:46 money. Last year, the world average was $108 and inside China, 84. China now makes over 80% of the cells. The shape repeats. Do not try to win the market at its current price. Remove the price,

5:59 absorb the losses, and end up owning the layer underneath once the competition can no longer afford to stay in the room. Software compounds harder than steel does because a mill cost a billion

6:09 dollars and a fork costs nothing. Anyone can take those weights, tune them, break them, publish the fix, and the next release inherits every one of those experiments for free. Linux is the

6:20 version of this argument that already finished. A student project now runs all 500 of the world's 500 fastest supercomputers. Selling Linux itself is a tiny business. The money moved to

6:30 everything sitting on top of it. Which brings us back to the model because a strategy argument is easy to make when the product goes unchecked. The launch post says it is competitive with much

6:41 larger models. Read that twice. It is doing careful work. The chart underneath is doing something less careful and the chart is what traveled. Dot studio published two comparison charts. 20

6:52 groups of bars, 12 on reasoning and agents, eight on multimodal. Their model is the first bar in every single group and the only bar painted a different color. In all 20 of those groups, it is

7:03 never the tallest bar. The appendix behind the charts is more useful. Anyway, 26 benchmarks, 12 models, and their model tops exactly two rows. One is an instruction following test called

7:14 ifbench at 80.4, a real win. It also loses on both of the benchmarks dots studio wrote itself. The second win is ARC AGI3 on the official harness where the winning score is 6.9 out of 100. It

7:29 wins because the two Frontier models tested there scored 1.5 and 0.4. Here is the number that should stay with you. The same model on the same benchmark scores 32.1 when it runs through a

7:41 general harness instead of the official one. same weight, more than four and a half times the score purely from the scaffolding around them. The headline that traveled was Arc AGI 2 81.4 against

7:54 61.4 for Deepseek's V4 Flash, the July checkpoint. 20 points from a model with 16 billion active parameters. An observer posting as Toraxes flagged it just under 4 hours before the launch

8:06 post because the weights had sat on hugging face for days. The post said it loses to the July checkpoint on quote mundane evaluations. Then added three words carrying the sentence if real.

8:18 Footnote 2 tells you why that hedge was there. Their 81.4 was measured on the public evaluation set by them. The rival figure is Deepseek's official score from the private set run by the people who

8:30 own the benchmark and that one is independently verified at exactly that number while DOT 3 has not appeared on that leaderboard at all. So here is the call in two halves pointing opposite

8:41 ways as a model to actually run. This is not the one. 577 GB and 8 GPU node even compressed and 393 downloads in the week since it appeared. It sits 10 points behind Claude Opus 4.8 on Swebench

8:57 verified and 13 behind Kimmy K3 on Terminal Bench. For nearly every viewer here, Quen or DeepSeek is still the correct download this week and it is not close. As a strategy, it already won.

9:09 And the model being merely good is the point rather than the flaw. A lifestyle app can now build something in this class inhouse and hand it out as a defensive expense. The teams it actually

9:19 serves are the ones building agents that run for hours. Credit where it is owed because this correction only exists thanks to Dot Studio. They printed the footnotes, named the harness, marked

9:29 their own runs with an asterisk, and blocked hugging face during testing so the models could not look up answers. The bar chart oversells. The document underneath it does not. One last thing

9:39 complicates it. On the 21st of July, China's Commerce Ministry met Alibaba, Bite Dance, and the lab behind GLM to discuss restricting foreign downloads of Chinese weights, filings for weak

9:51 models, security reviews for strong ones, a possible ban for the strongest. 24 days later, Red Note published the full model under Apache 2.0. At the World AI conference in Shanghai that

10:03 same month, Xiinping told the room to seize this rare historic opportunity to encourage open source, then called for monitoring and early warning in the same speech. So, the closing question is not

10:14 whether commoditization works. It plainly does. It is this. If the model layer is the cheap one now, which layer is the expensive one? And has anybody in this story stopped to check who already

10:24 owns