Kimi K3: China Hit the AI Frontier — and Gave It Away
The model that broke both walls America was counting on to stay ahead of China in AI.
Executive summary
For three years, America's lead in artificial intelligence rested on two walls. The first was physical: a widening embargo on the most advanced chips, meant to starve Chinese labs of the compute that frontier models are built on. The second was cultural: a safety-first instinct at the leading US labs — red-teaming, refusals, alignment research, weights kept locked behind an API. The bet was that hardware plus caution would keep the best models American, and keep them supervised.
In one week in July 2026, a three-year-old Beijing lab that most people outside China cannot name walked through both walls. Moonshot AI released Kimi K3, and on the independent Western leaderboard that frontier labs actually cite, it landed inside the top pack — statistically level with Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5, a short step behind the two absolute leaders. It was built by a team cut off from the best Nvidia chips. It was handed out as a free, downloadable file that anyone on Earth can run on their own machines. And in the detail that should worry Washington most, it was not cheap: Kimi K3 is priced almost exactly like a premium American model, not the rock-bottom bargain that defined the last wave of Chinese AI.
That combination is the story. The much-discussed gap between the US frontier and China's best has collapsed from something like 18 months to a matter of months. The chips were supposed to prevent this, and did not. The open weights mean the US can restrict its own closed models but has no lever at all over a file already mirrored across the internet — which is why the American government is now openly debating whether it can ban a piece of software it cannot delete. And the man who built Kimi K3 says, on the record, that alignment is just a problem you solve by scaling, that interpretability may be unnecessary, and that "AGI will be inherently global." This is not a story about one model. It is the week the terms of the US–China AI contest visibly changed.
Why one Chinese model release is a turning point
The AI revolution moves faster than any single person can track; the useful skill is knowing which events actually reset the board and which are noise. Most model releases are noise — a benchmark nudged, a price shaved. A few are load-bearing, because they retire a belief that a lot of strategy was resting on. For eighteen months the load-bearing Western belief was that China was a "fast follower": always a year or two behind, able to copy cheaply but not to lead, held back by chip controls and reliant on distilling US models. Kimi K3 is the event that makes that belief hard to keep. To understand why, you have to look past the headline number to three things underneath it — what the model actually is, how good it really is, and what it costs — and then at the people and the politics that produced it. That is the order this takes them in, because those are the questions whose answers changed.
Kimi K3 is a 2.8-trillion-parameter model that thinks by default — and is free to download
Start with what it is, because the shape of the machine explains most of what follows. Kimi K3 is a mixture-of-experts model with 2.8 trillion total parameters — but for any given word it processes, it lights up only 16 of its 896 internal "experts," roughly 1.8% of the network. That sparsity is the whole trick: the model has the knowledge capacity of something enormous while the cost of running it stays closer to a model a fraction of its nominal size. It reads up to a million tokens of context at once — a small library in a single prompt — and it is natively multimodal, understanding images as well as text.
Two design choices separate it from the American frontier less by capability than by philosophy. First, it "thinks" by default: where earlier models shipped a separate reasoning variant you had to switch on, K3 reasons through problems as standard, with an adjustable effort dial. Second, and decisively, its weights are open. Moonshot put K3 live in its app and API on 16 July 2026 and committed to releasing the full model weights — the actual trained file — by 27 July, under a permissive, MIT-style licence. Anyone can download it, run it on their own hardware, inspect it, fine-tune it, and build on it without asking permission from Moonshot or anyone else.
The technical differentiator underneath the sparsity is an attention mechanism Moonshot built and named Kimi Delta Attention — the machinery that lets the model attend across a million tokens without the cost curve exploding, which is the specific bottleneck that has kept most Western models at context windows a fraction of that size. And the "thinking" is tunable rather than binary: developers set an effort level — low, high, or max — trading latency and token cost against depth of reasoning, so the same model can answer a quick question cheaply or grind through a hard one deliberately.
The other thing worth grasping is the cadence, because it explains why this arrived faster than anyone expected. K3 is not a lab's first serious attempt; it is the top of a staircase climbed at a sprint. Kimi K1.5, in January 2025, already claimed to match OpenAI's o1 on maths and coding. Kimi K2, in July 2025, was a one-trillion-parameter mixture-of-experts model with 32 billion active parameters, open-sourced and pitched squarely at agentic work. From there the releases came almost monthly — K2.5 in January 2026, K2.6 in April, a coding-specialised K2.7 in June — before K3 crossed the three-trillion-parameter line in July. Five significant releases in twelve months is not the rhythm of a follower catching its breath; it is the rhythm of a lab that has found its stride.
So the similarities are real and the differences matter. Under the hood, K3 is the same species of machine as the models you already use — we pulled apart what actually separates the leading Western models in our comparison of GPT-5.5 and Claude a couple of months back, in episode 8, and K3 belongs in that same conversation. What is different is everything about how it reaches you: not a metered service you rent through a company's servers, but a file you own.
On the test frontier labs actually cite, it sits inside the top pack
The claim that matters is not "it's good for a Chinese model." It is that on neutral ground, judged by the same yardstick the US labs use in their own marketing, Kimi K3 is a frontier model. Artificial Analysis — the independent benchmarking outfit whose composite "Intelligence Index" has become the closest thing the field has to a shared scoreboard — placed K3 third overall on release, with a score of about 57 across its blend of reasoning, coding, and knowledge tests. That put it essentially level with Claude Opus 4.8 and GPT-5.5, and a step behind only the two current leaders, Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol.
The composite average even understates the case in the places that pay. On agentic tasks — the long, tool-using, multi-step work that enterprises actually deploy models for — K3 ranks at or near the top of Artificial Analysis's GDPval agentic index, ahead of GPT-5.5 and Claude Opus 4.8 on that measure, and it takes first place outright on the AutomationBench test of end-to-end task completion. On at least one front-end coding arena it ranks first as well, and independent testers noted its cybersecurity performance sits "extremely close" to GPT-5.6 Sol's, at a fraction of the price. The one place it looks ordinary is factual reliability: on the Omniscience test that pairs accuracy against hallucination, its numbers are middling — a reminder that "frontier-class" is not "flawless," and, as we will see, a detail that matters a great deal to how its makers think about safety.
This is not a model that wins a narrow benchmark and falls apart in use; it is broadly, genuinely capable, and it is the first open-weight model to force its way into a group that has been exclusively American and closed. The belief that the frontier is a US monopoly did not survive the week.
The real tell isn't the benchmark — it's that Kimi stopped being cheap
Here is the detail that separates a curiosity from a turning point, and it is the one almost everyone missed in the rush to compare benchmark scores. The last wave of Chinese models — DeepSeek above all — shocked the world on price: frontier-adjacent quality at a tenth or a twentieth of the cost, the "super cheap Chinese AI" that made Western labs' margins look suddenly fragile. Kimi K3 broke that pattern deliberately. It is expensive.
Kimi K3 costs $3.00 per million input tokens and $15.00 per million output tokens. Its predecessor, K2.6, released only three months earlier, cost $0.95 and $4.00. The output price alone nearly quadrupled. And $3 in, $15 out is not a Chinese price — it is precisely what Anthropic charges for Claude Sonnet 5, a model K3 outperforms. We dug into exactly why a "cheaper" frontier model can still command a premium in our Claude Sonnet 5 episode a couple of weeks ago, episode 24; Moonshot has just applied the same logic in reverse, charging a premium because the model earns it.
Why would a Chinese lab give up its one famous advantage? Because pricing is a signal, and Moonshot is sending one: we are not the budget option anymore, we are a peer. Undercutting is what challengers do when they cannot compete on quality; pricing at parity is what you do when you can. It also quietly reframes the whole "cheap Chinese AI will destroy US margins" panic. The threat is no longer a race to the bottom on price — it is a Chinese model good enough to charge American prices while being free to self-host for anyone who would rather not pay at all. That is a harder problem, not an easier one.
It comes from a three-year-old Beijing lab most Westerners can't name
The company is Moonshot AI — 月之暗面 in Chinese, "the dark side of the moon," a nod to Pink Floyd. It was founded in March 2023, which means the model that just cracked the global frontier was built by an organisation younger than the iPhone 15. Its headquarters are in Beijing, and its chatbot, Kimi, is one of China's most-used AI assistants.
The people are the point. Moonshot's chief executive, Yang Zhilin, is a computer scientist born in Shantou in 1992. He studied at Tsinghua University, then took a PhD at Carnegie Mellon under Ruslan Salakhutdinov — one of deep learning's most cited researchers — where he co-authored Transformer-XL and XLNet, two genuinely foundational papers on how models handle long sequences of text. He worked at Google Brain and Meta's AI lab before returning to China to build Moonshot with two Tsinghua classmates. Zhou Xinyu specialises in deploying neural networks on constrained hardware — precisely the skill a chip-starved lab would prize — after stints at Hulu and Tencent. Wu Yuxin, now Moonshot's chief operating officer and head of AI research, spent years at Meta's FAIR lab, where he co-developed Group Normalization and built Detectron2, two pieces of infrastructure that a large slice of the computer-vision field ran on. This is not a team that learned AI by copying OpenAI. It helped write the research the whole field is built on, then went home.
The money has followed the pedigree at a startling pace. Moonshot raised at roughly an $18 billion valuation in March 2026, then about $20 billion in May as demand for its open models surged, and is now reportedly closing a round near $31.5 billion — with pre-IPO discussions said to be starting as early as August 2026. Alibaba and Tencent, China's two consumer-internet giants, sit among the backers, alongside venture firms like IDG Capital. Yang describes the company's DNA as a fusion of "OpenAI's technology idealism with the business philosophy of ByteDance" — the research ambition of a frontier lab welded to the ruthless product-and-growth machine of the company behind TikTok. Chinese domestic press, for its part, has settled on a simpler label for the young founder at the top: the country's "baby-faced billionaire."
It caught up fast because the chip embargo forced it to get efficient
How does a lab locked out of the best hardware ship a frontier model? The uncomfortable answer for the architects of the export controls is: partly because of them. Denied the top Nvidia chips, Chinese labs could not win by brute force — throwing ever more compute at the problem, the strategy that has defined much of the US frontier. They had to win on efficiency, and scarcity is a ferocious teacher.
That is why Kimi K3 activates under 2% of its parameters per token, and why Moonshot built a custom attention mechanism it calls Kimi Delta Attention to handle a million-token context without the cost exploding. The company claims K3 achieves roughly 2.5 times the overall scaling efficiency of its predecessor — more capability per unit of compute, which is exactly the axis a compute-constrained lab is forced to optimise. Add three more ingredients: the elite returnee talent already described; an open-weights strategy that turns the entire global developer community into an unpaid testing and improvement flywheel; and the proof-of-concept that DeepSeek delivered in early 2025, which showed the whole Chinese ecosystem that a lean lab could rattle the US frontier. Stack those together and the timeline compresses hard.
The DeepSeek release of early 2025 was the hinge. It did not just embarrass the assumption that China was hopelessly behind; it handed every other Chinese lab a proof of concept and a playbook — that a lean team, denied the best hardware, could rattle the US frontier and command the world's attention doing it. Moonshot is what that proof looks like industrialised: not a one-off shock, but a lab shipping frontier-adjacent models on a monthly cadence. The open-weights choice compounds the advantage, because releasing the model turns millions of outside developers into an unpaid extension of the lab — finding bugs, building tooling, fine-tuning variants, publishing benchmarks — a diffusion flywheel a closed lab simply cannot spin as fast.
The analyst Nathan Lambert, who tracks the open-model ecosystem closely, put it plainly: the gap between the American closed frontier and China's best open models has shrunk from a debated six-to-nine months to something more like three-to-five. And crucially, he argues, this is no longer explained away by distillation — the idea that Chinese labs are just cheaply copying US models. "Chinese companies," he wrote, "are extremely good at building models in the same way the leading American companies are." The catch-up is real capability, not a shortcut. Beijing appears to read it the same way: Xi Jinping's public embrace of open-source AI reads less like a hope than a bet already paying out.
The chip wall was supposed to stop exactly this — and it didn't
Step back to the strategy. The core American bet, sustained across two administrations, was that controlling the supply of advanced AI chips would keep Chinese models a safe distance behind. The logic was clean: frontier models need enormous compute; compute needs the best chips; deny the chips and you deny the frontier. Kimi K3 is a large, loud data point against that logic.
The Brookings Institution, examining the DeepSeek shock, had already concluded that export controls were showing their limits — Chinese labs kept finding ways to optimise around the hardware constraint faster than the constraint could bite. There are persistent, disputed reports that some Chinese labs have obtained restricted Nvidia chips despite the controls, and separate US claims that banned Blackwell-class hardware found its way into Chinese training runs; treat those as contested rather than settled.
Meanwhile the constraint is quietly loosening from the other side, as China builds its own silicon. Huawei's Ascend line has moved from a curiosity to a workhorse: a research group including Huawei recently completed full post-training of DeepSeek's 1.6-trillion-parameter V4-Pro on a cluster of Ascend 910C chips, and other Chinese labs have trained state-of-the-art models end-to-end on homegrown hardware. Ascend still lags Nvidia on raw efficiency, and there is no public confirmation of exactly what Kimi K3 trained on — but the strategic direction is unmistakable. Every year the embargo holds is a year China spends building a compute stack it fully controls.
But the deeper problem for the embargo is that it may be aiming at the wrong target. If scarcity breeds efficiency, and efficiency breeds frontier models at a fraction of the expected compute, then the chip wall does not just fail to stop Chinese AI — it actively trains Chinese labs to need fewer chips, and buys them time to build their own. We traced the scramble to break Nvidia's grip on AI silicon back in April, in our episode on Nvidia's challengers, number 6; the export-control fight is the geopolitical face of that same contest over who controls the compute.
And the open-weight wall can't be built at all
Even if the chip controls worked perfectly, they would run into a second, harder problem — one with no hardware fix. The United States can restrict its own frontier models, because they are closed: OpenAI and Anthropic serve them through APIs, behind terms of service, subject to US law. Washington has no equivalent lever over an open-weight model. Once Moonshot posts the Kimi K3 file, it propagates across Hugging Face mirrors, torrents, and private servers worldwide within a couple of days. There is no server to shut down, no company to subpoena, no update to withhold. The file simply exists, everywhere, forever.
And giving the weights away is not charity — it is strategy, and a shrewd one. A US Congressional commission review of China's open-model push, aptly titled "Two Loops," argued that open-sourcing frontier models reinforces China's industrial dominance: every developer who builds on Kimi or DeepSeek is one more user of a Chinese software stack, one more voice pulling global standards, tooling, and habits toward Chinese defaults rather than American ones. Closed models earn margin on each call; open models earn something a challenger values more — ubiquity, and a claim on the ecosystem everyone else has to interoperate with. When you are behind on paid market share, making the frontier free is how you leapfrog it.
This is why the policy response has turned slightly surreal. Catalysed directly by Kimi K3's release, the US administration is reportedly weighing restrictions on the use of cutting-edge Chinese AI models inside the country — a debate about whether you can, in effect, ban a free file that is already on developers' laptops. By May 2026, open-weight models — the great majority of them Chinese — accounted for roughly 61% of the tokens flowing through open systems, with four of the five most-used open models coming from China. When the open ecosystem the world's developers build on is largely Chinese, "just block it" stops being a coherent policy; it is closing the barn door on an animal that has already been photocopied. This is the sovereignty question we opened up in our Sovereign AI episode, number 17, from last month, arriving from the opposite direction: not America's power to switch others off, but its growing inability to switch anything off.
Wall Street had a DeepSeek flashback — then bought the dip
None of this happened quietly in markets, and it is worth being precise about what actually occurred, because the headlines oversold it. When K3 landed, traders reached instantly for the DeepSeek template. On 18 July 2026, semiconductor and AI stocks sold off: Taiwanese equities fell more than 6% intraday, Japanese names around 4%, and the Nasdaq slipped roughly 1.5%. The fear was familiar — if a cheap, capable open model can do the work, why keep pouring hundreds of billions into proprietary infrastructure and the chips beneath it?
But the panic was shallow and brief. Most of those losses were recovered within the same session; investors, burned into learning from the original DeepSeek scare — which also faded fast and proved a buying opportunity — declined to repeat it. And the smarter analysis noted a real distinction: DeepSeek's story was about making models radically cheaper to train and run, whereas Kimi K3, as Bloomberg observed, is a very large model that leans heavily on memory infrastructure — arguably more demand for certain hardware, not less. The honest read is that K3 is a genuine strategic event dressed, for a day, in the costume of a market event. The thing to watch is not the stock chart. It is the policy and the capability underneath.
Moonshot's idea of safety is the mirror image of Anthropic's
Now to the question that should unsettle anyone paying attention to AI risk: how does the lab building these models think about safety? The answer is that Moonshot's philosophy is close to the opposite of the safety-first US labs' — and that difference is becoming a competitive variable, not just a moral one.
Yang Zhilin has been unusually direct. Where Anthropic has built its identity on interpretability — the painstaking work of understanding a model's internals so you can trust and steer it — Yang has argued the reverse: "If a model never hallucinates or has an extremely low hallucination rate, interpretability wouldn't even be necessary." His view of alignment is that it is not a distinct hard problem requiring a distinct hard discipline, but something scaling itself resolves: "Alignment itself also follows a scaling law, meaning it can be solved as long as the right data can be found." In this worldview, you do not inspect the black box; you make it good enough that you do not have to.
This is not just a debate about virtue. An independent safety evaluation of the earlier Kimi K2.5 found it roughly as dual-use-capable as GPT-5.2 and Claude Opus 4.5 — but with significantly fewer refusals on requests touching chemical, biological, and weapons-related harms, and more compliance with disinformation and copyright-infringement prompts. Because the weights are open, whatever guardrails ship with K3 can be fine-tuned away by anyone who downloads it. And that raises a genuinely destabilising possibility, one the South China Morning Post put bluntly: that the US labs' safety caution — the refusals, the red lines, the deliberate slowing — may be handing China a speed advantage. That fear is exactly what our episode this month on Musk, Amodei, and Karp, number 29, was circling: the worry that warning loudly about the danger might be a luxury the winner cannot afford. Kimi K3 turns that worry from an abstraction into a data point.
Its real target isn't a chatbot — it's AGI, and it thinks AGI has no flag
Finally, the deepest question: what does Moonshot actually want? Not, on its own account, a successful chatbot business. Yang has framed Moonshot's singular mission as achieving artificial general intelligence, with commercial products explicitly described as intermediate steps and byproducts of that goal — a concentration of every resource on one target that he contrasts with hedging across many. His stated path there is idiosyncratic and worth understanding: he believes long context is the road to AGI. He traces the history of AI as a steady expansion of how much context a model can hold — from a handful of words to hundreds of thousands of tokens — and makes the striking claim that "if you have a billion-token context length, then the problems we face today would no longer be problems." A model that can hold enough of the world in mind at once, in his telling, starts to look general.
He pairs this with a prediction about compute that is quietly radical: that by 2026, the computation a model spends on inference — thinking, and generating its own training data — may far exceed what it spent on training, a paradigm he sums up as "inference is training." Read literally, it means the bottleneck stops being how much human data you can scrape and becomes how much thinking the model can afford to do about itself; a model gets smarter by running, not just by being built. K3's always-on reasoning, with its dial that lets it spend more compute per answer, is that thesis made product — and if he is right, it also blunts the sharpest edge of the chip embargo, since a lab that can turn inference into training needs less of the frontier hardware that training was supposed to require. We went deep on the safety scientists mapping the road to superintelligence — Yoshua Bengio among them — last month in episode 14; Yang is running at the same destination from an almost opposite temperament, treating the path as an engineering problem to be scaled through rather than a risk to be gated.
And then there is the line that should stay with policymakers longest. Asked how he thinks about AGI in a divided world, Yang has said he "firmly believes AGI will be inherently global." From the CEO of a Beijing lab, cut off by American export controls, giving his frontier model away to the entire planet, that is not a throwaway. It is a worldview — that the most powerful technology in history will not respect the walls nations are trying to build around it. Whether that is idealism or strategy, the release of Kimi K3 is what it looks like in practice.
Bottom line
The lasting significance of Kimi K3 is not a leaderboard position that the next US release will reclaim. It is the retirement of a comfortable assumption. The idea that America could hold a durable AI lead through hardware control and a head start now has to survive the fact that a chip-starved, three-year-old Beijing lab reached the frontier anyway, gave the result away, and charged full price for the convenience of not having to run it yourself.
Two questions now define the contest, and neither is the one the export controls were designed to answer. The first is open versus closed: if the world's open-model ecosystem is overwhelmingly Chinese, and open weights cannot be recalled, then a growing share of what everyone builds on will be downstream of Beijing's release schedule — and no American restriction reaches it. The second is speed versus caution: if capability and safety really do trade off, then the US labs' interpretability-and-restraint culture is either the responsible path or a self-imposed handicap, and Kimi K3 is the strongest evidence yet that the market may not reward the careful.
Watch three hinges. Whether China's domestic chips can sustain this pace without smuggled hardware — if they can, the embargo is spent. Whether any US move to restrict open models proves even faintly enforceable — if it doesn't, the open-weight wall is confirmed unbuildable. And whether Yang's core bet holds — that you can scale your way to a model so capable it is safe by default. Because if alignment really is just a scaling law, the lab that scales fastest wins, and it has stopped being obvious that lab is American. The wall isn't the chips. It never was. It was the belief that being ahead was something you could keep.
Sources - Kimi K3 quickstart and technical documentation — Moonshot's own specs: 2.8T parameters, 16-of-896 experts, 1M context, thinking modes, open-weights date. - Kimi K3 achieves #3 on the Artificial Analysis Intelligence Index — the independent benchmark placing K3 level with Opus 4.8 and GPT-5.5. - Kimi's open model K3 signals the end of super-cheap Chinese AI — the pricing shift: $3/$15, near-quadruple K2.6, at parity with Claude Sonnet 5. - Kimi K3: the open-weights escalation (Nathan Lambert, Interconnects) — the gap narrowing from 6-9 to 3-5 months and the "genuine capability, not distillation" argument. - Kimi (chatbot) — Wikipedia) — the model timeline from the 2023 launch through K2, K2.5, K2.6 and K3, with parameter and licence details. - Yang Zhilin — Wikipedia — the founder's background: Tsinghua, Carnegie Mellon, Transformer-XL, XLNet, Google Brain and Meta. - Moonshot AI's AGI vision (ChinaTalk) — Yang on long context as the road to AGI, "inference is training," alignment as a scaling law, and interpretability. - China's Kimi K3 fuels fears safety curbs are holding back US AI (SCMP) — the safety-versus-speed argument and K3's near-parity cybersecurity performance. - Kimi K3 and the open-weight enforceability problem (FourWeekMBA) — why the US cannot gate an open-weight file and the restriction debate it triggered. - Kimi K3 triggered DeepSeek flashbacks for the stock market (Yahoo Finance) — the 18 July sell-off and rapid recovery. - DeepSeek shows the limits of US export controls on AI chips (Brookings) — the case that chip controls train Chinese labs to need fewer chips.
Transcript
Sam: For three years, America had two walls keeping it ahead of China in AI. This week, a three-year-old lab in Beijing walked straight through both of them.
Alex: And here's the part that should keep people up at night — it built a frontier model with second-rate chips, gave it away for free to the entire planet, and then charged full American prices for the version you don't have to run yourself.
Sam: Free and premium at the same time. That shouldn't be possible.
Alex: It shouldn't. And that contradiction is the whole story.
Sam: Welcome back to Dan's AI Intel — the show where we try to make sense of the fastest, most consequential shift most of us are ever going to live through.
Alex: The idea here is pretty simple. This stuff is moving faster than any one person can track, the jargon multiplies every week, and most of it is noise. So we try to do the opposite of noise — take the handful of things that actually move the board, and get you both the breadth and the real depth of what happened, in one sitting, so you come away genuinely understanding it.
Sam: I'm Sam, that's Alex, and today we are getting into a model called Kimi K3.
Alex: Which, if you don't follow this world closely, is a name you have almost certainly never heard. And that is exactly why it matters.
Sam: So here's the setup. A week or so ago, a Chinese lab most Westerners couldn't pick out of a lineup released an AI model that landed, on the independent scoreboard the big American labs actually cite in their own marketing, inside the top pack — level with the best that OpenAI and Anthropic have shipped.
Alex: And I want to be clear up front why that's a genuine event and not just another benchmark headline. For eighteen months the settled Western belief was that China was a "fast follower" — always a year or two behind, good at copying cheaply, but not able to lead. A lot of strategy was quietly resting on that belief being true.
Sam: And this is the release that makes it hard to keep believing.
Alex: This is the one that retires it. That's what makes it load-bearing rather than noise — it knocks out an assumption a lot of other thinking was standing on. And the reason I couldn't put this one down isn't the leaderboard position — those change month to month. It's what this release lets us ask. For years the entire American strategy rested on two beliefs: that you could keep the best models American by controlling the chips, and keep them safe by keeping them locked up and cautious. This is the week both of those beliefs got tested in public.
Sam: So we're going to travel through a few places. What this thing actually is, under the hood. How good it really is, on neutral ground. The one number almost everyone missed — which is the price. Then the people who built it, and how a lab that was supposed to be starved of chips caught up this fast. And finally the part that gets genuinely unsettling: how these people think about safety, and about where all of this is going.
Alex: And I'll flag it now — there's a turn near the end, a thing the man who built this says out loud, that flips the whole export-control strategy on its head. We'll get there. But if you sit with what it implies, it's hard to un-hear.
Sam: Before we dig in — if you get anything out of this, the single most useful thing you can do is follow the show. Whatever app you're in right now, there's a follow or a plus button, it's one tap and it's free, and it's genuinely the biggest thing that helps a small independent show like this keep going.
Alex: So let's start where you have to start — with what this machine actually is. Because the shape of it explains almost everything that comes after.
Sam: Okay. Kimi K3. Tell me what I'm actually looking at.
Alex: So on paper it's enormous. Two-point-eight trillion parameters. Parameters are basically the model's adjustable knobs — the things it tunes during training to store what it knows. Two-point-eight trillion of them makes this one of the largest models anyone has ever released.
Sam: Two-point-eight trillion. I have no intuition for that number at all.
Alex: Nobody does, and you don't need one — because here's the trick. For any given word it's processing, it does not switch all of those on. It's what's called a mixture-of-experts model. Think of it as a huge building full of specialists — eight hundred and ninety-six of them. And for each word that comes through, a little router at the door taps just sixteen of them on the shoulder and says, you handle this one.
Sam: Sixteen out of eight hundred and ninety-six. So it's this vast body of knowledge, but it only ever wakes up a tiny slice of it at a time.
Alex: Under two percent at a time. And that's the point. You get the knowledge capacity of something gigantic, but the cost of actually running it stays closer to a model a fraction of the size. It's how you build something this big without the bill being insane.
Sam: That's clever. It's like having a thousand-person law firm on retainer but only ever paying the two people who touch your case.
Alex: That's exactly it. And there are two more things about it that matter. First, it "thinks" by default. Older models, you had to flip a switch to get them to reason step by step — there'd be a separate reasoning mode. This one reasons through problems as standard, and it's got a dial: developers can set the effort to low, high, or max, trading speed and cost against how hard it grinds on a problem.
Sam: So cheap and quick for "what's the capital of France," crank it to max for something genuinely hard.
Alex: Right. Same model, tunable depth. And then the second thing, which is the one that changes everything downstream — the weights are open.
Sam: Okay, "the weights are open" is one of those phrases people say like it's obvious. Pretend it isn't. What does that actually mean?
Alex: Fair. The weights are the trained model itself — the actual finished file, the thing that is the intelligence. Most of the American frontier models, you never touch that file. It sits on OpenAI's or Anthropic's servers, and you rent access through what's called an API — you send a question over the internet, the answer comes back, but the model stays locked in their building.
Sam: I'm renting a car I never get to keep.
Alex: You're renting a car you're not even allowed to look under the hood of. Open weights is the opposite. Moonshot — the company — put K3 live in its app in mid-July, and committed to releasing the entire trained file to the public a couple of weeks later, under a very permissive licence. Which means anyone on Earth can download it, run it on their own machines, take it apart, retune it, and build on it — without asking Moonshot for permission, and without Moonshot ever knowing.
Sam: So it's not a car I rent. It's a car anyone can just... have. Forever. And copy.
Alex: Now you've got it. And hold onto that, because that single fact — it's a file, not a service — is the thing Washington cannot figure out what to do about. But we'll build to that.
Sam: There was a technical bit you mentioned — a special attention thing?
Alex: Kimi Delta Attention. Don't worry about the name. The problem it solves is this: these models have a "context window," which is how much they can hold in mind at once. K3 can hold up to a million tokens — call it a small library — in a single prompt. The catch is that normally, the more you ask a model to hold in its head, the more the cost explodes, kind of exponentially. Delta Attention is the piece of engineering Moonshot built to let it read that whole library without the cost curve blowing up.
Sam: And most Western models can't hold nearly that much.
Alex: Most sit at a fraction of it. And this matters more than it sounds, because the man who built this thinks a giant context window isn't a feature — he thinks it's the road to something much bigger. That's a thread we're going to pull hard at the end.
Sam: You keep teasing the end. I see what you're doing.
Alex: I make no apologies. One last thing on what it is — the speed of it, because it's the part that genuinely spooked people. K3 didn't come out of nowhere. Back at the start of 2025, they had a model — K1.5 — that already claimed to match OpenAI's o1 on maths and coding. Mid-2025, K2 — a one-trillion-parameter version with thirty-two billion of those parameters active per word, open-sourced, and aimed squarely at that agentic, tool-using work.
Sam: So even a year ago they were at a trillion parameters and giving it away.
Alex: And then they just kept shipping. A version in January, another in April, a coding-specialised one in June — roughly monthly — until K3 crossed the three-trillion line in July. Five serious releases in twelve months.
Sam: That's not somebody catching their breath. That's somebody who's found a rhythm.
Alex: That's a lab that's found its stride. And a monthly cadence is its own kind of statement — it says we're not going to surprise you once and disappear. We're going to be here every few weeks. Which brings us to the obvious question — is it actually any good, or is this a case of big numbers and a good press release?
Sam: Right, because "large" and "good" are not the same thing. I could build a giant model that's confidently wrong about everything.
Alex: Completely. So let's be careful about what the claim is. The claim is not "it's good for a Chinese model" — that's the faint-praise version everyone's used to. The claim is that on neutral ground, judged by the exact yardstick the American labs use in their own marketing, this is a frontier model, full stop.
Sam: And who's holding the yardstick? Because if it's Moonshot's own benchmark, I don't care.
Alex: Good instinct, and no. There's an independent outfit called Artificial Analysis. They run a composite they call the Intelligence Index — it blends reasoning, coding, and knowledge tests into one number, and it's become the closest thing this field has to a shared scoreboard. When K3 came out, they put it third overall. A score of about fifty-seven.
Sam: Third. Out of everyone. Who's ahead of it?
Alex: Only the two absolute newest American flagships — Anthropic's Claude Fable 5, and OpenAI's GPT-5.6 Sol. And they're ahead by a whisker — sixty and fifty-nine, to K3's fifty-seven. And critically, K3 lands level with Claude Opus 4.8 and GPT-5.5, which until about ten minutes ago were the models everyone held up as the state of the art.
Sam: So it's not sniffing at the frontier from behind. It's standing in the group.
Alex: It's in the room. And it's the only open model in that room. Everything else up there is closed, American, and rented.
Sam: Now, I've learned to be suspicious of a single average, because averages hide things. Is there anywhere it's actually weak?
Alex: Great question, and yes — and the honest answer cuts both ways. Where it's freakishly strong is the stuff businesses actually pay for. The long, multi-step, tool-using work — "agentic" tasks, they're called, where the model has to go off and do a whole job across many steps, not just answer a single question.
Sam: Give me a concrete version of that, because "agentic" is one of those words that's started to mean nothing.
Alex: Sure — think "book me a trip": check the calendar, search flights, compare prices, fill in the form, handle the error when the first card fails. Many steps, using real tools, without a human holding its hand at each one. That's the work enterprises actually deploy these for. And on the independent index that measures exactly that, K3 ranks at or near the very top — ahead of GPT-5.5 and Opus 4.8. On one test of finishing a task end to end, it takes first place outright. On a front-end coding arena, first as well. And on cybersecurity tasks, independent testers put it extremely close to GPT-5.6 Sol — one of the two best American models there is — at a fraction of the price.
Sam: Okay so where's the weakness?
Alex: Factual reliability. There's a test that pairs raw accuracy against how much a model just makes things up — hallucinates — and there K3 is middling. Not embarrassing, but ordinary.
Sam: Huh. So it's a brilliant doer that's a little loose with the facts.
Alex: That's a fair caricature. And hold that thought, genuinely, because that specific weakness — the hallucination number — turns out to connect directly to how these people think about safety. It's not a random flaw. It's downstream of a philosophy. We'll close that loop.
Sam: Another one for the pile. Fine. But the headline stands — this is a real, broadly capable model.
Alex: It's the first open-weight model to force its way into a club that has been exclusively American and exclusively closed. And by the way — if you want the deeper anatomy of what actually separates the top Western models from each other, we pulled GPT-5.5 and Claude apart in episode 8, a couple of months back. Quick aside, but K3 belongs squarely in that same conversation now.
Sam: So the monopoly's gone. But you keep hinting the benchmark isn't even the real story.
Alex: It isn't. The real story is a number almost nobody clocked, because they were all busy comparing test scores. Let's talk about the price.
Sam: The price. Okay. My assumption — and I think everyone's assumption — is that the whole point of Chinese AI is that it's cheap. That's the DeepSeek story from last year, right? Frontier-ish quality for pennies.
Alex: That's exactly the assumption. And it's the assumption K3 detonates on purpose. Because the last wave — DeepSeek above all — shocked everyone on price. Something close to frontier quality at a tenth, a twentieth of the cost. The "super cheap Chinese AI" that suddenly made Western labs' profit margins look fragile.
Sam: Right. So K3 undercuts them all again.
Alex: No. That's the twist. K3 is expensive.
Sam: Wait. Deliberately expensive?
Alex: Deliberately expensive. Here are the numbers. K3 costs three dollars per million words of input, and fifteen dollars per million words of output — output being what the model generates back to you. Its own predecessor, from three months earlier, cost ninety-five cents in and four dollars out.
Sam: So in one release they nearly quadrupled the output price. On their own product.
Alex: Nearly four-x. And here's the line that should make an American lab executive put their coffee down. Three in, fifteen out is not a Chinese price. It is, to the dollar, exactly what Anthropic charges for Claude Sonnet 5 — a model that K3 outperforms.
Sam: Okay, that genuinely stops me. Why on earth would you give up the one advantage everyone knows you for? Being the cheap option is a real weapon.
Alex: It is. And the answer is that price is a message, and Moonshot is sending one very deliberately. Undercutting is what you do when you can't compete on quality — you compete on the receipt instead. Pricing at parity is what you do when you can compete on quality, and you want everyone to know it. It's the difference between a brand that's always on sale and a brand that never needs to be.
Sam: It's the luxury move. Charging full price isn't about the money. It's the signal that says, we're not the discount aisle, we're the real thing.
Alex: That's precisely it. The price tag is the flex. And funnily enough, we dug into exactly this counterintuitive thing — how a "cheaper" model can still command a premium — in our Claude Sonnet 5 episode a couple of weeks ago, episode 24. Quick aside: Moonshot has just run that same logic in reverse. Charge the premium, because the model earns it.
Sam: But there's a sting in this, isn't there? Because it quietly kills the comforting story Western labs were telling themselves.
Alex: It does, and this is the part I'd underline. The comforting story was: sure, Chinese models are cheap, but cheap means low-margin, low-margin means they can't fund the next leap, so we're fine. K3 erases that. Now the threat isn't a race to the bottom on price. It's a Chinese model good enough to charge American prices — while also being free to download and run yourself if you'd rather not pay at all.
Sam: Oh, that's brutal. So it's premium and free at the same time, depending on what you want.
Alex: Premium if you want the convenience of someone else running it. Free if you've got the hardware and don't. And that is a much harder problem to compete against than "they're cheaper." You can beat cheap. What do you do about "as good as you, at your price, and also free"?
Sam: And it changes the whole shape of the threat, doesn't it. The old worry was margins — cheap Chinese AI dragging everyone's prices down until nobody can fund the next model.
Alex: That was the exact fear after DeepSeek — a race to the bottom that quietly bankrupts the people trying to push the frontier. K3 says: that's not the threat anymore. The threat is a peer that's chosen not to undercut you, because it doesn't need to. It would rather take the margin and the respect. Undercutting was the tell that they knew they were second-best. Charging parity is the tell that they've stopped believing that.
Sam: It's almost a confidence thing as much as a business thing.
Alex: It's confidence made into a price tag. And confidence is contagious — it's the thing that pulls in the next round of funding, the next wave of talent, the next set of developers who want to bet on the winner. The price isn't just revenue. It's recruitment.
Sam: Yeah, I don't have an answer to that, and I don't think they do either. So — who are these people? Who looks at the board and makes that move?
Alex: That is the right question. And the answer is not who you'd guess. The company's called Moonshot AI. Though the Chinese name is better — it translates as "the dark side of the moon."
Sam: That's a Pink Floyd reference.
Alex: It's a straight Pink Floyd nod, yes. And the fact worth sitting with is the age. Moonshot was founded in March 2023. So the lab that just cracked the global AI frontier is younger than the iPhone 15.
Sam: Come on. Three years old? OpenAI's been at this for a decade.
Alex: Three years old. Beijing-based. They make one of China's most-used AI assistants — the chatbot's called Kimi. But the story isn't really the company. It's the people. Because the easy assumption in the West is that Chinese labs got here by copying OpenAI. These founders helped write the papers OpenAI is built on.
Sam: Okay, back that up. Who's the person at the top?
Alex: Yang Zhilin. Computer scientist, born in 1992. Studied at Tsinghua — think China's MIT — then did his PhD at Carnegie Mellon under Ruslan Salakhutdinov, who is one of the most cited researchers in all of deep learning. And while he was there, Yang co-authored two papers — Transformer-XL and XLNet — that are genuinely foundational to how modern models handle long stretches of text.
Sam: So this isn't a guy who read about AI. He helped invent pieces of it.
Alex: And Salakhutdinov isn't a footnote there — he's one of the founding figures of modern deep learning. So Yang learned this straight from the people who invented it. He then worked at Google Brain and at Meta's AI lab before going home to start Moonshot. And he didn't go alone. One co-founder, Zhou Xinyu, specialises in getting neural networks to run on limited hardware — which, if you're a lab that's about to be cut off from the best chips, is possibly the single most valuable skill on Earth.
Sam: That's almost too on the nose. The chip ban is coming, and they've got the guy whose whole thing is making AI run on less.
Alex: It's remarkable, isn't it. And the third one, Wu Yuxin — now their chief operating officer and head of AI research — spent years at Meta's FAIR lab building things like Group Normalization and a tool called Detectron2 that a huge chunk of the computer-vision world literally ran on top of. So this isn't one star founder with a supporting cast. It's three people who each built infrastructure the field still leans on.
Sam: So the picture is — these are returnees. Elite training, worked inside the American giants, then took all of it back to Beijing.
Alex: That's the picture, and it's the thing that undoes the "fast follower" story. This is not a team that learned AI by copying it. They helped write the field's foundational research, and then they went home to build.
Sam: And the money found them, I assume.
Alex: Startlingly fast. They raised at roughly an eighteen-billion-dollar valuation in March of this year. About twenty billion in May — as demand for their open models surged. And they're reportedly now closing a round near thirty-one and a half billion, with talk of an IPO as early as August. Alibaba and Tencent — China's two consumer-internet giants — are among the backers, alongside big venture firms.
Sam: So they nearly doubled the valuation in a couple of months.
Alex: In a couple of months. That's not a company being valued on today's revenue — nobody's revenue grows like that. It's a company being valued on where people think the frontier is going, and who they think is going to be standing on it.
Sam: Those valuations are moving faster than the models are.
Alex: They kind of are. And Yang has this line about the company's DNA I keep chewing on. He describes it as a fusion of, quote, "OpenAI's technology idealism with the business philosophy of ByteDance." ByteDance being the company behind TikTok.
Sam: So — moonshot research ambition, welded to the most ruthless growth-and-product machine on the planet.
Alex: That's the self-description. The research idealism of a frontier lab, bolted to the go-to-market instincts of TikTok. And the Chinese press has landed on a nickname for him that tells you how young he is — they call him the country's "baby-faced billionaire."
Sam: A baby-faced billionaire who helped invent the Transformer's memory and now runs a thirty-billion-dollar lab. Okay. But talent and money don't fully explain the speed. There's a wall in this story — the chips. How did a lab that's supposed to be starved of hardware move this fast?
Alex: And this is where it gets genuinely uncomfortable for the people who built the wall. Because the honest answer is: partly because of the wall.
Sam: Say that again, because it sounds backwards. The chip ban helped them?
Alex: It sounds backwards and it's the crux of the whole thing. So — the setup. For years, America's plan to stay ahead was to control the chips. Frontier models need enormous amounts of computing power. That power comes from the best chips, which are Nvidia's. So the logic was: deny China the top chips, and you deny them the frontier. Clean, simple, and it's held across two US administrations.
Sam: And it makes total sense on paper. Starve them of the fuel.
Alex: On paper. But here's what actually happened when you cut a brilliant, motivated lab off from brute force. They couldn't win by just throwing more compute at the problem — the way a lot of the American frontier does, honestly. So they were forced to win on efficiency instead. And scarcity is a ferocious teacher.
Sam: This is the classic thing where the constraint makes you more creative. Give someone unlimited budget and they get lazy. Cap it hard and they get ingenious.
Alex: That's the mechanism exactly. And you can see it in the design. Remember the "only sixteen of eight hundred experts fire" trick — activating under two percent of the model per word? That's an efficiency move. The custom attention engine so it can read a million words without the cost exploding — efficiency move. Moonshot claims K3 gets roughly two-and-a-half times the overall efficiency of its predecessor — meaning way more capability squeezed out of each unit of computing power.
Sam: Which is exactly the muscle you'd build if you didn't have enough chips to be wasteful.
Alex: It's the only muscle you can build. And you stack a few things on top. The elite returnee talent we just talked about. An open-weights strategy that basically turns millions of outside developers into an unpaid army testing and improving your model for free. And a proof of concept — DeepSeek, in early 2025 — which showed the entire Chinese ecosystem that a lean team, denied the best hardware, could genuinely rattle the American frontier and grab the world's attention doing it.
Sam: So DeepSeek wasn't just a one-off shock. It was a template.
Alex: It was the hinge. Early 2025, DeepSeek didn't just embarrass the assumption that China was hopelessly behind — it handed every other Chinese lab a proof of concept and a playbook. That a lean team, denied the best hardware, could genuinely rattle the American frontier and command the world's attention doing it. And Moonshot is what that playbook looks like industrialised — not a single lucky punch, but a lab shipping frontier-adjacent models on a monthly beat.
Sam: And the open-weights thing feeds that loop too, doesn't it? Because it's not just marketing.
Alex: It's the accelerant. When you release the model to the world, you turn millions of outside developers into an unpaid extension of your own lab. They find your bugs, they build your tooling, they fine-tune their own variants, they publish benchmarks you didn't have to run. A closed lab, keeping everything in-house, simply cannot spin that flywheel as fast — because it's doing all of that work with only the people on its payroll.
Sam: Is there a number on how much the gap has actually closed? Because "they're catching up" is vague.
Alex: There is, and it's from someone worth trusting on it — an analyst named Nathan Lambert, who tracks the open-model world about as closely as anyone. His read: the gap between the American closed frontier and China's best open models has shrunk from a debated six-to-nine months down to something more like three-to-five.
Sam: From, what, a year and a half a couple of years ago, down to a single-digit number of months.
Alex: And here's his sharpest point. He argues this is no longer explainable as distillation — that's the dismissive theory that Chinese labs are just cheaply copying the outputs of American models. His line was: "Chinese companies are extremely good at building models in the same way the leading American companies are." In other words, it's real capability now, not a shortcut.
Sam: And I imagine Beijing sees it the same way.
Alex: Xi Jinping has publicly embraced open-source AI as a strategy. And when you look at K3, that embrace reads less like a hope and more like a bet that's already paying out. By the way — the wider scramble to break Nvidia's grip on the chips underneath all this, that was our episode 6, on Nvidia's challengers, back around April. Quick aside: the export-control fight is really just the geopolitical face of that same contest over who controls the compute.
Sam: Okay so if the chip wall is leaking — let's actually stand in front of it. Was it just a bad idea from the start?
Alex: Not a bad idea. But maybe a wall built facing the wrong direction. Let me lay out why. So the core bet, again, was: control the advanced chips, keep Chinese models a safe distance behind. K3 is a large, loud data point against that bet. And it's worth being fair about the evidence, because this is where people overclaim in both directions.
Sam: Good, because my instinct is to go straight to "the ban failed, tear it down," and that's probably too fast.
Alex: It is too fast. So — a few honest pieces. One: a serious analysis from the Brookings Institution, looking at the earlier DeepSeek shock, had already concluded the export controls were hitting their limits, because Chinese labs kept optimising around the hardware constraint faster than the constraint could actually bite.
Sam: So the smart people flagged this before K3.
Alex: They flagged the pattern — that Chinese labs kept finding ways to optimise around the hardware limit faster than the limit could actually bite. Two: there are persistent but disputed reports that some Chinese labs have gotten hold of restricted Nvidia chips anyway, and separate American claims that banned high-end hardware found its way into Chinese training runs. I want to be clear — treat those as contested, not settled. Nobody's publicly confirmed exactly what K3 was trained on.
Sam: And that matters, because it's the easy out, isn't it. "They only caught up because they cheated and got our chips."
Alex: It's the comforting version, and it might even be partly true. But it can't be the whole story — because the efficiency gains are real and measurable, and they show up in the architecture whether or not there were smuggled chips in the building. If your entire explanation is "they cheated," you don't have to change your strategy. If the real explanation is "the constraint made them better," you have a much bigger problem, and you'd rather not look at it.
Sam: Right, so we can't just say "they cheated with smuggled chips" and move on.
Alex: We can't, and it'd be lazy. But here's the third piece, which is the one that actually matters. The constraint is quietly loosening from the other side — because China is building its own chips. Huawei has a line called Ascend, and it's gone from a curiosity to a genuine workhorse. A research group including Huawei recently did the full final-stage training of DeepSeek's 1.6-trillion-parameter model on a cluster of homegrown Ascend chips.
Sam: So not just "we found a workaround for Nvidia." More like "we're growing our own Nvidia."
Alex: That's the direction, and a 1.6-trillion-parameter model is not a toy — that's a serious training run finished end to end on homegrown silicon. Ascend still lags Nvidia on raw efficiency — this isn't parity yet, and I don't want to overstate it. But the trajectory is unmistakable. Every year the embargo holds is a year China spends building a compute stack it fully controls. And here's the deeper problem for the whole embargo, the thing that makes it almost self-defeating. If scarcity breeds efficiency, and efficiency breeds frontier models on a fraction of the expected compute — then the chip wall doesn't just fail to stop Chinese AI.
Sam: It trains them to need fewer chips.
Alex: It actively trains them to need fewer chips. And it buys them the time to go build their own. Which is close to the worst outcome the policy could produce.
Sam: That's the part that flips it from "the ban isn't working" to "the ban might be backfiring." You wanted to make them dependent, and instead you taught them to be independent.
Alex: You made them fit by putting them on a diet. And even if — big if — the chip controls worked perfectly, they run headfirst into a second wall. A wall that has no hardware fix at all. And honestly, this second one is the one I find harder to argue with.
Sam: A second wall. Okay. What's the first thing that breaks about controlling this stuff once the model is open?
Alex: Everything, basically. Here's the asymmetry. The United States can restrict its own frontier models — because they're closed. OpenAI and Anthropic serve them through an API, behind terms of service, subject to US law. There's a company, a server, a lever you can pull.
Sam: There's a somebody to call.
Alex: There's a somebody to call. But once Moonshot posts the K3 file to the internet, within a couple of days it's mirrored everywhere — copied onto servers, torrented, sitting on developers' laptops all over the world. There's no server to shut down. No company to subpoena. No update you can withhold. The file just exists. Everywhere. Permanently.
Sam: It's the thing about the internet everyone learns the hard way. Once something's out, it's out. You cannot un-publish a file that's already been copied ten thousand times.
Alex: It's the leaked-song problem, at the scale of the most powerful technology of the era. You can shut down the original uploader all day. It does nothing. The copies are the point.
Sam: But hold on — why would a company give away the thing it spent all that money building? That's the bit I never understand about open-weights. Isn't that just torching your own asset?
Alex: It looks like charity and it absolutely isn't — it's strategy, and a shrewd one. There was a US Congressional commission review of China's open-model push, and it had this great title — "Two Loops." And the argument was that open-sourcing your frontier model reinforces China's industrial dominance. Because every single developer who builds on Kimi is now one more user of a Chinese software stack. One more voice pulling the world's tools and standards and habits toward Chinese defaults instead of American ones.
Sam: So it's a bit like the old line about standards — whoever's foundation everyone builds on quietly sets the rules for everyone else.
Alex: That's the game exactly. It's the platform play. If the world's developers learn AI on your model, format their data for your model, write their tools around your model — you don't need to win any single sale. You've won the layer underneath all the sales.
Sam: So the free model is the loss leader, and the prize is — everyone building on your foundation.
Alex: The prize is ubiquity. Closed models earn margin on every call. Open models earn something a challenger values more — they earn the ecosystem. They earn being the thing everyone else has to interoperate with. And Yang's logic is basically: when you're behind on paid market share, you don't try to out-sell the leader. You make the frontier free and leapfrog the whole game.
Sam: How much of the open world is already Chinese, then? Is this theoretical or is it already happening?
Alex: It's already happened. As of about May this year, open-weight models — the big majority of them Chinese — accounted for roughly sixty-one percent of the usage flowing through open AI systems. Four of the five most-used open models in the world came from China.
Sam: Sixty-one percent. So the open ecosystem isn't drifting Chinese. It basically is Chinese.
Alex: Which is why the policy response has gotten a little surreal. Triggered directly by K3's release, the US administration is reportedly weighing restrictions on using cutting-edge Chinese AI models inside the country. Which means Washington is now seriously debating whether it can ban a free file that is already sitting on American developers' laptops.
Sam: You can't. I mean — can you? How do you ban a file people already have?
Alex: You can't, cleanly. That's the point. Think about what enforcement would even look like. You'd have to police what software runs on private laptops and company servers across the whole country. The file's already downloaded, already mirrored, already fine-tuned into a dozen variants with different names. There's no upload to take down, because the upload happened days ago and the copies don't need it anymore.
Sam: It's less "ban a product" and more "ban a number." Once the file exists, it's just... math that's out in the world.
Alex: That's exactly the right way to see it. And "just block it" stops being a coherent policy the moment the thing you're blocking has already been photocopied a million times. And I want to connect this to something, because it's the exact mirror of a story we told last month. In episode 17, Sovereign AI, we looked at America's power to switch other countries off — to deny access to its AI like a kill switch. Quick aside: this is that same sovereignty question arriving from the opposite direction. Not America's power to switch others off — its growing inability to switch anything off.
Sam: Ooh. That's a real turn. The superpower that built the kill switch is discovering there's a whole class of thing it can't switch.
Alex: That's the discomfort in one sentence. Now — none of this happened in silence. When K3 landed, the markets had a full-body flashback. And that's worth getting right, because the headlines oversold it.
Sam: Right, I saw "markets tank on Chinese AI" headlines. What actually happened?
Alex: Let's be precise, because the precise version is more interesting than the headline. On the eighteenth of July, when K3 landed, chip and AI stocks sold off. Taiwanese stocks fell more than six percent at the worst point in the day. Japanese names around four percent. The Nasdaq slipped roughly one and a half percent.
Sam: And the fear driving that is — what, exactly?
Alex: The exact same fear as the original DeepSeek scare. It goes: if a cheap, capable, open model can do the work, then why are American companies pouring hundreds of billions into proprietary infrastructure and the chips underneath it? Maybe all that spending is about to look foolish.
Sam: So investors reached for the DeepSeek template and hit sell.
Alex: They reached for the template. But here's the thing — the panic was shallow and it was brief. Most of those losses got recovered within the same trading session.
Sam: Same day? So the fear evaporated almost as fast as it arrived.
Alex: Because investors had been burned into learning. The original DeepSeek scare, last year, also cratered things — and then faded fast and turned out to be a buying opportunity. So this time a lot of them just... declined to panic twice.
Sam: Fool me once. That's almost funny. They'd already run this exact drill.
Alex: They had. And the smarter analysis pointed out a real distinction between the two events. DeepSeek's whole story was about making models radically cheaper to train and run — less demand for expensive hardware. K3, as Bloomberg noted, is a very large model that leans heavily on memory infrastructure. Which is arguably more demand for certain hardware, not less.
Sam: So the reflex was "this is DeepSeek again, sell the chipmakers," but under the hood it might actually be good for some of them.
Alex: Which is why the honest read is this: K3 is a genuine strategic event that spent one day wearing the costume of a market event. The stock chart is the least interesting thing about it. The thing to actually watch is the policy and the capability underneath — which is everything we've been talking about.
Sam: It's funny how the market keeps making the same category error. It sees "capable Chinese model" and reaches straight for "sell the chipmakers," when the actual event is geopolitical, not financial.
Alex: Because a market can only price what it can model, and "this changes the balance of power between two superpowers over the next decade" doesn't fit in a one-day trade. So it gets squeezed into the nearest familiar shape — the DeepSeek shape — and then unwound by lunchtime. The mispricing isn't that they panicked. It's that they filed a strategic earthquake under "volatility" and moved on.
Sam: Okay. So we've done what it is, how good, the price, the people, the chips, the open-weight trap, the markets. There's one place left, and you've been dangling it the whole time. How do these people think about keeping this stuff safe?
Alex: And this, for me, is where it stops being a business story and starts being something that should genuinely unsettle you. Because their answer is close to the opposite of everything the cautious American labs believe.
Sam: The opposite how? Give me the shape of it.
Alex: So take Anthropic — the makers of Claude — as the archetype of the safety-first camp. Their whole identity is built on something called interpretability. That's the painstaking work of actually understanding what's happening inside the model — opening up the black box, mapping its internals — so you can trust it and steer it. The bet is: understand it first, then deploy it.
Sam: Understand the engine before you drive the car.
Alex: Exactly that. Now here's Yang Zhilin, and he's been unusually blunt. He's basically argued the reverse. His line — and I'm quoting — "If a model never hallucinates, or has an extremely low hallucination rate, interpretability wouldn't even be necessary."
Sam: Wait. So his view is — if the thing just doesn't make mistakes, who cares what's going on inside it?
Alex: That's the view. Don't bother opening the box. Just make the box good enough that you never need to. And remember that hallucination weakness we flagged way back — the one place K3 scored ordinary?
Sam: Oh. That's the loop closing, isn't it. Their safety philosophy is literally staked on the one thing the model isn't yet great at.
Alex: It's staked on getting the hallucination rate to basically zero. Their bet is that they can. And it goes further. On alignment — which is the whole discipline of making sure a powerful AI actually does what we want and shares our goals — Yang's position is that it isn't some distinct, fiendish problem needing its own field. His words: "Alignment itself also follows a scaling law, meaning it can be solved as long as the right data can be found."
Sam: Unpack "scaling law" for me, because that phrase is doing a lot of work.
Alex: A scaling law, in AI, is this observed pattern that models just predictably get better as you throw more data and more compute at them. More scale, more capability, like clockwork. What Yang is saying is: alignment is the same. It's not a special problem. Make the model big enough and feed it the right data, and it comes out aligned. Safety falls out of scale.
Sam: So one camp says safety is a hard, separate thing you have to carefully engineer and inspect. And the other says, nah — safety is just a side effect of making it powerful enough.
Alex: You've got the two worldviews exactly. And I want to be really clear this isn't just a philosophy-seminar difference, because it has teeth. There was an independent safety evaluation of the earlier Kimi model, K2.5. It found the model roughly as capable of dangerous dual-use stuff as GPT-5.2 and Claude Opus 4.5 — comparable capability. But — significantly fewer refusals on requests touching chemical, biological, and weapons-related harm. And more willingness to go along with disinformation and copyright-infringement requests.
Sam: So it's about as dangerous in what it knows, but much more willing to actually tell you. The knowledge is the same; the willingness to hand it over is the difference.
Alex: That's the distinction that matters. Capability and restraint are two different dials, and the American labs have been turning the restraint dial way up — refusing, red-lining, slowing themselves down on exactly the chemical, biological, and weapons stuff. The K2.5 evaluation suggests Moonshot's dial is set much looser. Same rough capability, fewer brakes. And now layer the open weights on top. Because the weights are public, whatever guardrails do ship with K3 can be fine-tuned right off by anyone who downloads it. You can just... sand the safety features away at home.
Sam: Right, because it's your file now. There's no company keeping the guardrails bolted on remotely. You own the whole thing, safety catches included.
Alex: And that's the bit that makes the open-versus-closed question stop being abstract. With a closed model, the lab can watch how it's being used, patch it, pull access if someone's abusing it. With an open one, the moment it's downloaded, all of that is gone. The safety story ends at the download button.
Sam: So even if Moonshot ships it perfectly cautious — which, to be clear, the evaluation says they don't — it wouldn't matter, because anyone can undo it.
Alex: It's belt-and-braces in reverse. The one place you can enforce safety — keeping the model on your own servers — is the exact thing they've given up on principle. And that raises a genuinely destabilising possibility — the South China Morning Post put it bluntly. The worry is that the American labs' caution — the refusals, the red lines, the deliberate slowing down — might be handing China a speed advantage. That being careful is, in a race, a handicap.
Sam: That's a horrible thought. That doing the responsible thing is how you lose.
Alex: It's the thought that's been haunting this whole field, and it's exactly what we were circling in our episode this month on Musk, Amodei, and Karp — episode 29. Quick aside: the fear that warning loudly about the danger might be a luxury the eventual winner just can't afford. K3 takes that fear and turns it from an abstraction into a live data point.
Sam: Okay but let me push back a little, because "they don't refuse as much" could just be a less polished product, not a worldview.
Alex: It's a fair challenge, and on its own you'd be right to hold it loosely. But it stops being ambiguous when the CEO tells you the worldview out loud — repeatedly, on the record. This is a philosophy, not an oversight. And it points at the last, deepest thing, which is: what does Moonshot actually want? Because it is not, on their own account, to build a nice chatbot business.
Sam: Not a chatbot business. So what's the actual goal?
Alex: AGI. Artificial general intelligence — very roughly, an AI that's as broadly capable as a human being across basically anything, rather than being good at one narrow slot. Yang has framed Moonshot's entire mission as achieving AGI, and he describes the commercial products — the chatbot, the API, all of it — explicitly as intermediate steps. Byproducts on the way to the real target.
Sam: So Kimi the assistant, the thing millions of people use — that's not the point. It's scaffolding.
Alex: It's scaffolding for the actual project. And his path there is genuinely idiosyncratic, and worth getting, because it links back to something from the very start of this. He believes long context is the road to AGI.
Sam: Long context — that's the "hold a whole library in mind at once" thing from earlier.
Alex: The million-token window, yes. And here's his reasoning. He tells the history of AI as basically one long expansion of how much context a model can hold — from a handful of words, up to hundreds of thousands of tokens now. And he makes this striking claim: "If you have a billion-token context length, then the problems we face today would no longer be problems."
Sam: A billion. We're at a million and that's already a small library. A billion is — what, a whole field of knowledge held in mind at once?
Alex: A thousand times what K3 does today. His intuition is that a model that can hold enough of the world in its head simultaneously — every relevant fact, every prior step, the whole situation — starts to look general. Starts to look like it understands, rather than retrieves. Whether he's right is a genuinely open question, and plenty of researchers think it's only one piece of the puzzle — but that's his bet, and K3's giant context window is that bet made into a product you can use today.
Sam: So the million-token window isn't just a nice feature to him. It's a down payment on a theory of how you get to AGI at all. You said there was a line near the end that flips the chip strategy. Is this it?
Alex: This is it. He pairs the context idea with a prediction about compute that's quietly radical. He thinks that, starting around now, the computing a model spends on inference — that's the thinking it does when it's actually running, answering you, reasoning — may come to far exceed what it spent on training in the first place. He sums it up in three words: "inference is training."
Sam: Okay, walk me through why that's a bombshell and not just a slogan.
Alex: So — the old picture. Training is the giant, expensive one-time build: you pour in oceans of human data and mountains of compute, and out comes a finished model. Then inference — running it — is the cheap afterthought. Yang is saying that flips. The model gets smarter mainly by running — by thinking, and generating its own training data about itself — rather than by being built once from human data.
Sam: It's the difference between studying and practising. You can only read so many textbooks. But you can practise forever, and get better by doing.
Alex: That's a genuinely good way to put it. It learns by doing, not just by being taught. And there's a reason that's such a big deal — the old worry was that we'd run out of human text to train on. The whole internet is only so big; at some point you've read all of it. If the model can improve by thinking rather than by reading more of us, that ceiling disappears.
Sam: So the fuel stops being human writing and starts being the model's own thinking time.
Alex: Exactly. And K3's always-on reasoning, with that dial that lets it spend more compute per answer — that is this thesis turned into a shipping product. But now — here's the flip you asked for. If inference is where the intelligence increasingly comes from, then a lab that masters turning inference into training needs less of the frontier hardware that training was supposed to require.
Sam: Oh. So the chip ban was built to choke off training. And if the game is quietly moving from training to inference —
Alex: — then the ban is aimed at the wrong phase. It's guarding the front door while the value moves to the back. That's why I said sit with it. The whole export-control strategy assumes training is the bottleneck. Yang is betting the bottleneck is moving somewhere the ban doesn't reach.
Sam: And by the way — that whole question of how you actually get to superintelligence safely, whether it's even the right thing to chase — that's ground we covered.
Alex: It is — we went deep on the safety scientists mapping that road, Yoshua Bengio among them, in episode 14, a few weeks back. Quick aside: and the contrast is stark. Bengio's camp treats the path to superintelligence as a risk to be carefully gated. Yang is running at the same destination from almost the opposite temperament — as an engineering problem to be scaled straight through.
Sam: And then there's the actual line, isn't there. The one about a divided world.
Alex: Asked how he thinks about AGI when the world is splitting into rival blocs, Yang has said he "fir…