31. “26Q1”: Has the China-U.S. Foundation-Model Gap Widened or Narrowed Over the Past Year? — Hugging Face | 王铁震

Summary

• 王铁震’s bottom line: On actual usage, U.S. frontier models remain “far ahead,” for 3 structural reasons. Abundant compute lets U.S. labs ship monthly updates, while Chinese models may update once every 3 months; DeepSeek V4’s 1.6T model may be only “Flash or Nano level” by comparison; and closed-source development creates one-way transparency: “The U.S. can learn from China’s top exploration, while China cannot learn from the U.S.”
• One underappreciated advantage of DeepSeek V4 is moving KV cache from memory to SSD, potentially creating a 10x price difference between a cache hit and a miss. 王铁震 says DeepSeek is “doing public service”: it is investing heavily in architectural exploration and actively adapting to Huawei; the V4 launch page says in the fine print that prices will fall again once Huawei’s 950 POD cluster comes online. He expects the next Kimi and GLM models to borrow the approach, “making intelligence cheaper in China”—a major contribution.
• Cards are a constraint, but valuation and monetization also determine training scale. 王铁震 notes that 智谱 is valued at RMB400B versus Anthropic’s $1T-plus valuation, while MiniMax and 智谱 each generated roughly $100M in revenue over the past 12 months; even if B300 exports were allowed, they could not afford many GPUs. “They particularly need VC investors to fund them at 100x or 1,000x valuations and burn money to build a future.” Raymond later summarized that Chinese models “may” have only 1% of their peers’ training resources and revenue; willingness to pay has already been proven—“智谱’s monthly subscriptions are sold out.”
• China’s global share of open-source models is severely underestimated: HF puts it at 41% versus 36.5% for the U.S., excluding mainland Chinese users who cannot access HF. U.S. developers went through 3 stages—from the Reddit viral post “Qwen is great, but I have no way to use it,” to Qwen replacing Llama as researchers’ default base model, to major tech companies using “open-source model” to mean a Chinese model 80% of the time. Cursor used Kimi to train Composer 2.
• Closed-source data flywheels are the U.S.’s moat, but a 1.2B-person flywheel may be China’s answer. Raymond believes Chinese models only need to cross the usability threshold to accelerate inside a domestic flywheel outside the “isolated island” formed by U.S. models, much as Douyin did. His switching condition is specific: reach Claude 4.6 quality, and he would switch completely.
• One-way learning has run in the opposite direction too: R1 was the first open-source model to teach the world how to implement chain-of-thought. In the o1 era, nobody knew how chain-of-thought was achieved and some guessed at Monte Carlo search; 王铁震 says it is “very possible” Anthropic used DeepSeek R1 to train its first chain-of-thought model because DeepSeek got there earlier.
• 王铁震 expects the 2026 wave of movement-style open source to ebb: open source should be a go-to-market tool, not a technical ideal. If a company already holds 80% of the open-source market, he says there is little reason to keep investing at the same pace; once companies start making money, incentives change, and monetization is what makes open source sustainable. Stability’s “too-late focus on making money—the CEO had already left” is the counterexample. Citing MiniMax’s 闫俊杰, he calls open source “the core tool for building a company’s advanced-technology brand.”
• On Hassabis’s “6 months behind plus a mentality problem” argument, 王铁震 says the two sides are looking at different things: “one is 王健林, the other is 杨振宁.” China lacks the U.S.’s tolerance for the kind of trial and error that reportedly lets Thinking Machines Lab raise $5B before producing a product; everyone is under pressure to monetize. But electric vehicles and solar show that late movers can overtake incumbents, while DeepSeek’s architectural innovation leaves him “still fairly optimistic.”

Deep dive

1. The bottom line: U.S. frontier models feel “far ahead”

• 王铁震 looks at usability, not parameter counts: “We vote with our feet.” The primary tools are Opus and Codex. The first reason is compute: once B300 comes online, U.S. models may move to monthly iteration while Chinese models update once every 3 months, “quietly opening up a gap of several months.”
• The second is model size. Some say DeepSeek V4’s 1.6T parameters are too large to run inference on, but U.S. models are far larger: “For them, 1.6T may just be Flash or Nano level.” In a market where “bigger is better,” it is hard to compare 1.6T with 6T.
• The third is one-way transparency: “We don’t know what technology OpenAI used and have no way to learn from Anthropic’s experience,” while the architectural exploration represented by DeepSeek is fully open source. “The U.S. can learn from China’s top exploration, while China cannot learn from the U.S.”

2. The usability threshold: Raymond pushes back on the optimistic narrative

• 王铁震’s relative measure is a line between usable and unusable. When he held the “world’s first large-model novel-writing competition” in 2023, nobody used a domestic model; today, 智谱’s models are being snapped up by the market, and Chinese models are “rapidly crossing” the line at which they can generate revenue and get users to pay.
• Raymond’s pushback is worth preserving: he uses every model, but for day-to-day work only Claude Code is usable—“Codex is unusable.” He once ran the “Little Lobster” project on MiniMax, but it could not handle multi-step tasks or calls to unusual tools: “If it isn’t intelligent enough to explore, the task cannot be completed. That is unusable.”
• 王铁震’s reconciliation is that most Chinese users still interact with models through chat windows, while major domestic tech companies are rapidly requiring employees to become AI-native. The two usability thresholds are measuring different user groups to begin with.

3. DeepSeek V4’s splash—and the underappreciated KV cache: “It’s doing public service”

• Why was V4 less explosive than the R1 moment? The disillusionment curve explains it. On first seeing ChatGPT 3.5, 王铁震 thought, “I can retire now.” After 4.6, users find 4.7 underwhelming if it fails to surprise them: “Normally, our progress is linear, but people have exponential expectations.” The market is brutal: “If you can’t keep surprising me, you’re not good enough.”
• But 王铁震 believes V4 has an advantage that has not been fully recognized: it compresses KV cache from a large footprint to a small one. “KV cache used to have to sit in memory; now it can sit on an SSD.” That means storing more for longer and still hitting the cache when a conversation resumes the next day. A hit versus a miss could mean a 10x price difference; combined with DeepSeek’s “price-cutter” pricing, coding becomes very cheap.
• 王铁震 acknowledges that he has not used 智谱 and MiniMax particularly seriously, so his judgment may be colored by past experience.
• His characterization is that DeepSeek is one of the few Chinese companies capable of large-scale exploration of foundation-model architecture. Kimi and GLM are trying to imitate its structure, and he expects the next Kimi or 智谱 model to borrow from it too, “making intelligence cheaper in China.” Token economics matter to the whole world.

4. Domestic-chip adaptation is a global-scale problem—and DeepSeek is actually doing it

• The difficulty is that no one else in the world matches NVIDIA. AMD and Intel have spent heavily but still cannot catch its ecosystem. 昇腾 is “more like an FPGA”: different operator implementations could make a model output ABC on NVIDIA GPUs and DEF on Huawei GPUs. “Then you can’t hand that model to users.” The entire stack—from libraries to frameworks to quantization—has to be rebuilt, after decades of NVIDIA development.
• Everyone is working on it, and Alibaba even has its own GPU, but DeepSeek has gone furthest down the Huawei path and achieved the best quality. The fine print on the V4 launch page says prices will fall again once Huawei’s 950 POD cluster comes online. “That shows a very deep tie-up with Huawei. It isn’t a matter of getting it to run and then issuing a PR release.”
• 王铁震 does not care what foreign media think: before V4 launched, countless outlets confidently said it would arrive the following week, and none did. What matters is the industry’s internal response—the approach to cutting inference costs will spread across the sector.

5. GPUs constrain training, but valuation and revenue set the scale too

• 王铁震 sees GPU availability as a ceiling on training, but not the only variable. Training is a one-time investment, while inference costs depend on usage and can theoretically be recovered through fees. Which is larger depends on inference volume.
• 智谱 is already valued at RMB400B, while Anthropic is a $1T-plus company. Even if Chinese companies were allowed to buy B300s, their smaller valuations and funding pools would prevent them from purchasing as many GPUs as U.S. companies.
• MiniMax and 智谱 each generated around $100M over the past 12 months. Even if they spent all of that revenue on GPUs, they could not buy many cards. “They particularly need VC investors to fund these companies at 100x or 1,000x valuations and burn money to build a future.”
• Willingness to pay is not the issue. Users rushed to buy 豆包’s monthly plan, Kimi’s monthly plan sold well, and “智谱’s monthly subscriptions are sold out.” “People are willing to pay, but you have to give me a model that reaches usable quality.”

6. A defense of benchmark chasing: The students are not at fault; no one set good problems

• 王铁震’s defense—“I am China’s student”—is that benchmark chasing is fundamentally a training-system and data-flywheel problem. After returning to China, 钱学森 asked students open-ended questions such as what specifications were needed to launch a rocket to the moon. Chinese models had not yet reached usable quality and did not know what users wanted. “Nobody had given them a good question, so they could only benchmark against the questions researchers commonly asked.”
• U.S. models finished the old benchmarks long ago and are “inventing new benchmarks to chase.” Those tests are built from massive user and enterprise demand and better reflect reality. “It is still benchmark chasing, but the benchmark is better.” Now that Chinese models have crossed the threshold and companies are setting problems, researchers have “suddenly found a treasure trove”—and after solving the benchmarks, they can still convert the work into real-world commercial applications.

7. Strange use cases and the closed-source data flywheel: “You’re paying to contribute decades of experience”

• Raymond’s hardest use case is video editing: more than 100 Claude Code conversations around a single video, calling Whisper for transcription, generating timestamps, determining “where the conversation begins,” and invoking image generation and editing functions. If any step fails, the video is unusable, and the user cannot intervene midway. “I’m sure Anthropic’s people probably never imagined someone would use a large model to make a video.”
• 王铁震 explains why closed source performs better. Subscriptions let vendors see countless cases like this and cover them in training in advance, producing the feeling that “this model understands me,” even though that understanding never appears on a benchmark. Once an open-source model is shipped, its creators do not receive user data and lack the closed-loop feedback.
• The sharper point is that user-written skills may be “mysteriously discovered by Anthropic and trained into the model” in the next version. “You’re not just paying to buy tokens; you’re actually paying to contribute decades of your experience.” Raymond adds that Harvey, a legal AI company with $100M in ARR, was replicated by one lawyer using Claude Code as a “massive skill” now downloadable by anyone on GitHub.

8. Will the gap keep widening? The 1.2B-person flywheel is the answer

• Raymond worries that with fewer Kimi subscribers and less data, the closed-source moat will simply keep getting higher. His view is that once Chinese models cross the usability threshold, they can develop outside the “isolated island” formed by U.S. models. 王铁震 agrees: “1.2B people can create an infinite amount of data for it.”
• The analogy is Douyin. Its video-recommendation algorithm was not invented by Douyin; 1.4B people helped spin its flywheel into the world’s best.
• Raymond’s switching condition is concrete. It is not patriotism; it is better Chinese-language support and the risk that a U.S. account could be shut down “in a matter of minutes,” taking out an important production tool. “If there were anything at 4.6-level quality, I’d be delighted and could switch over completely.”

9. Quant trading in practice: A strategy Cursor could not write a year or half a year ago now seems to work on 4.6

• Raymond cannot code at all and needs packaged, end-to-end delivery. About a year or half a year ago, he used Cursor to write a quantitative strategy, but “1 plus 1 sometimes equaled 3 and sometimes 4.” Once the context exceeded 200,000, the window collapsed completely. “When it gives you the feeling that it’s unusable, you absolutely cannot put money behind it in production.”
• Raymond now finds Claude 4.6 seems to have gotten somewhat smarter; when he rewrote those systems, “they seemed to work.” The key is skill: “A skill is essentially a way to compact context.”
• 王铁震 says some simple strategies made money, though not much. He recently used Claude Code to rewrite a complex CTA strategy and hopes it will make more this time. Could strategies stop working once everyone has access to large models? He thinks probably not, because strategy design involves aesthetic judgment. But he has not spent enough time on it to see a breakdown, nor built the patch that would follow one.

10. HF’s numbers and the 3 stages of U.S. developer adoption

• HF’s spring report put Chinese models at 41% of downloads versus 36.5% for U.S. models, but the report was “very conservative.” HF is inaccessible in mainland China, so Chinese downloads were excluded. The report covers global users but not mainland users. Chinese models are therefore severely undercounted, may be far above 41%, and had already overtaken U.S. models rather than doing so only when the report was published.
• Stage 1 was small businesses and individual developers. The tell was a Reddit post with thousands of replies: “Qwen is great, but I have no way to use it.” The author repeatedly had to explain to the company’s legal team, “This is just a stateless model.”
• Stage 2 was Qwen’s “large enough to eat your fill, with every size available” lineup and strong fine-tuning performance. It became researchers’ default base model and took Llama’s ecosystem position. When asked when Llama last updated, Raymond replied: “You suddenly mentioned that name and I realized I had forgotten it.”
• Stage 3 is that large U.S. tech companies “no longer care whether it’s a Chinese model.” When they say “open-source models,” they mean Chinese models 80% of the time. Llama is slow, IBM Granite is decent, and Google has only the small on-device Gemma models. Cursor used Kimi to train Composer 2. 王铁震 explains that 智谱 would have created a bigger issue because it is subject to export controls, so Cursor used Kimi instead: “At first, if it was Chinese, they didn’t want it. Later they stopped looking at China and looked only at open source.”

11. The other side of one-way learning: R1 taught the world chain-of-thought

• Raymond asks directly whether Anthropic is distilling Chinese models. 王铁震 says the current influence is more at the architectural level, but looking back to 2025, the chain-of-thought behavior of ChatGPT o1 was opaque. People guessed at Monte Carlo search. DeepSeek was the first to teach everyone through open source, after which every model converged aggressively in that direction.
• 王铁震 believes it is “very possible” Anthropic took DeepSeek R1 to train its first chain-of-thought model because DeepSeek got there earlier. In theory, Anthropic could also have run a dedicated internal distillation process on DeepSeek’s answers.
• A European aside: Mistral is doing well as France’s “sovereign large model” and is the obvious choice for politically aligned state-owned enterprises. German middle managers face far less resistance choosing Mistral than a Chinese model. But “unless you are truly irreplaceable,” BMW and Mercedes will still come to China to source autonomous-driving technology.

12. Prediction: 2026 may see a retreat from movement-style open source

• 王铁震 expects the post-2025 open-source frenzy—what he calls “movement-style open source”—to cool gradually in 2026. If a company already controls 80% of the open-source market, he says, there is little reason to keep investing at the same level. Once companies start making money, the views of bosses, investors, and shareholders all change: “Beyond flexing muscle, start making real money.”
• The right model is: “I hope these open-source companies make money aggressively and then open source aggressively, rather than selling everything they own just to open source.” Open source is a go-to-market, recruiting, and branding tool. Citing MiniMax’s 闫俊杰, 王铁震 says it is “the core tool for building a company’s advanced-technology brand”; top researchers do not want to join a closed-source company whose name will never appear in a technical blog or paper.
• Stability AI is the counterexample. It focused on monetization too late and completely lost the chance to continue open-sourcing; the company’s CEO left.

13. Responding to Hassabis: “One is 王健林, the other is 杨振宁”

• Hassabis said in January that China’s foundation models were about 6 months behind, had not shown the ability to push the frontier, and faced a mentality problem rather than a technology blockade. 王铁震’s framework is that Hassabis is speaking from the standpoint of someone trying to win a Nobel Prize: “You can look at the same thing and reach completely different conclusions.” A Nobel-level AI result from China is “quite difficult” because “there isn’t that much spare time to explore directions that may not work; everyone is too eager to make money.”
• But he is not pessimistic. Electric vehicles and solar are examples of late-mover success, and DeepSeek has shown that China can independently produce architectural innovation. “There is still hope for the future.” The U.S. can tolerate Thinking Machines Lab reportedly raising $5B without a product; China’s capital environment is less willing to give someone $5B to buy GPUs and explore without a clear reason. “This is a matter of time.”
• Raymond’s recap puts the sharpest number on it: the first line of the podcast episode on MiniMax’s IPO was, “It’s too hard.” Chinese models “may” have 1% of the resources of their counterparts, with training resources and revenue each also around 1%. 王铁震 embraces the charge of “blind optimism”: “Only when every investor is blindly optimistic can people in our industry be happier.”

14. A gap that is difficult to close soon: broad model safety and China’s top-down path

• 王铁震 cites articles and notes from Nathan Lambert and others who visited China. On broad model safety—how models affect everyone’s lives, rather than content controls—China “doesn’t care that much.” “Utility and doing good are not the same thing.” Examples include delivery workers “locked into algorithms” and 豆包’s approach of first giving users abundant emotional value; whether that is good or bad is unclear.
• Raymond observes a structural difference: China’s rules are largely backed by the government in a top-down process, while in the U.S. academic and industry leaders push legislation from the bottom up. 王铁震 adds the role of public opinion: Pinduoduo’s “cut one” campaign hit a wall in Europe. The next flashpoint could be youth mental-health problems from AI companionship, according to 王铁震; Raymond predicts “human burnout after coding tools”—“I feel like I’m a slave to AI.”

15. Two product-level non-consensus views: PPTs for office workers and the toxicity of “Code”

• 王铁震’s pitch is that once code penetration surges, the next immediately addressable but overlooked market is helping nontechnical office workers create PPTs and polished documents. Products like Claude Design use code to generate things nontechnical users can actually use. Raymond is already ahead of the curve: “Everything I write now is an HTML page that I print as a PPT.”
• Raymond’s advice to AI leaders at major tech companies is that the word “Code” is toxic. Many people around him can use all 3 modes but will not touch Code—not because it is difficult, but because “everyone outside programming has an innate fear of Code.” Code should be the underlying layer with 100% user penetration; users care only whether the final deliverable is delivered.

16. Hugging Face is the “Taobao of AI,” and the latest aha is model bootstrapping

• HF is a marketplace for models: vendors upload them, users access a unified API, and models can be downloaded or deployed in the cloud. Anyone who does not want to read papers can run a free demo in HF Spaces. The leaderboard reflects “researchers voting with their feet.” Raymond uses it to monitor domestic tech companies’ progress, while 王铁震 believes the length of time a model stays at the top can inform investment decisions.
• The past 90 days’ wow moment is getting models to iterate on themselves. AK’s AutoResearch project, released several months ago, gives a model a goal and lets it solve training problems on its own. “Writing code used to be the bottleneck, and Claude Code broke that. Then training models was the bottleneck; now large models are rapidly learning how to train models.”
• The end state is clear: “Once a model can fully bootstrap itself, you have effectively discovered a new species,” evolving and reproducing through its own equivalent of DNA. 王铁震 believes large Neo Labs may already have found something, and model self-iteration may already be happening without humans knowing. If the optimization target is wrong, it “may suddenly run very far,” creating a major ethical problem.

17. Epilogue: Raising children in the AI era—“Every generation will find its own mission”

• Reality is already rewriting the list of essential skills. 王铁震’s seven-year-old son starts typing by pressing the microphone and speaking; he can communicate without knowing how to write. 王铁震 himself uses Typeless coding and speaks from morning to night: “If everyone can use Typeless, why must children learn to write?”
• His reflection is that his parents once asked, “Will you be able to find a job if you play on the computer every day?” He did in fact find work by playing with computers, while people can now earn substantial money dancing in front of a camera. “Every generation has its own mission. Trust them—but what else can I do?”
• Raymond’s standard is relative ranking. Which university a child attends may not matter; what matters is “being the person among your peers who is more willing to try,” along with emotional stability. His alternative answer came from the pandemic: before it, he hoped his child would attend Peking University or Tsinghua; after it, the goal was simply to stay alive and avoid getting sick. 王铁震 adds that with or without AI, the most important things are physical health, doing what one wants to do, being happy, and maintaining a healthy mindset.