王铁震

Open source has caught up with the strongest closed models. 铁震’s review: the 1 year after DeepSeek V1 was ultimately “the boy-who-cried-wolf story”—closed models shipped monthly while open models shipped quarterly, and Opus 4.6, 4.7 and 4.8 remained far ahead; the first turn came with a model called GM 5.2, though even its name is disputed (he has already used it to “completely replace some of Opus 4.8’s functions”), followed by Kimi K3: “This isn’t about matching Opus on ratings—it has basically caught up with Fable, and in some scenarios it actually does better than Fable.” “Closed models suddenly realized they had no moat,” while MiniMax, Qwen’s 2-3T models and Zhipu are still on the way.

The distillation allegations do not hold up technically. Closed models expose neither logits nor chain-of-thought, making classic distillation impossible; Fable 5 launched in early July and K3 followed 15 days later—“there wasn’t even enough time to prepare the model release, let alone collect enough data.” Kis separates the claims into 3 layers: violating terms of service, systematic API scraping and stealing weights; the evidence points to “possible unauthorized large-scale data scraping,” but that does not mean K3’s core capabilities came from distillation. Distillation can get a model to a passing grade, but “K3 is already far beyond a passing grade”; focusing only on distillation is “missing the forest for the trees.”

Open-source inference is structurally cheaper, not subsidized. Closed API prices include a premium for the model itself, while Fireworks and Together AI charge only for hardware plus inference-service margin; K3 is more than 2x the size of K2, yet its scaling efficiency has improved 2.5x through technologies such as KDA, “mainly led by Chinese researchers.” Rumors put Fable’s base model at 8-10T parameters versus less than 3T for K3, which delivers comparable results; combined with infra optimization, “the cost of the same intelligence will fall another 1,000x over the next 3 years.”

K3’s license is a new template for monetizing open source. Model-as-a-Service providers with more than $20M in revenue over 12 consecutive months must sign a separate agreement with Moonshot AI before using K3 or derivatives in commercial services, with Vendor Verification built in. 铁震’s view: “A money-making open model is a good open model” (Stability’s demise is the warning), and cloud providers should welcome paying for official certification; if the model works, “this is pure money for nothing—you make the model good, and everyone simply pays tribute.” Kis warns that execution contains “a great many gray areas.”

The closed-model valuation framework is under pressure, just as the IPO year arrives. “OpenAI and Anthropic are both going public this year.” Monopoly power over intelligence was once “far more terrifying than monopolizing any other commodity” and therefore naturally worth trillions, but once intelligence becomes a liquid commodity, the margin story and the idea of funding unlimited spending under scaling laws become much harder to sell. Early signs of a price war are already visible: neoclouds are undercutting one another with open models, ChatGPT has started issuing coupons, and Anthropic’s account bans are “not nearly as crazy” as before. “It will all show up in their financial statements.”

The camps have split along business lines. The July 24 open letter on “Open Weights and American AI Leadership” grew from 25 signatories to roughly 75; it was 黄仁勋’s first post after joining X, while Anthropic declined to sign and Dario Amodei published a separate statement. Shovel sellers such as NVIDIA, whose CUDA moat was built in large part on open source, and companies with distribution such as Meta and Alibaba can embrace open source; OpenAI and Anthropic, which only sell APIs, have interests that “will never be aligned.” The agent application layer is being squeezed at the same time—one VC friend’s counterintuitive rule is: “The higher the valuation, the more willing we are to invest.”

The safety narrative has been turned upside down. K3 scored only in the 30% range on the cyber-offense and defense benchmark Exploit Bench, versus above 75 for frontier models; after Hugging Face was attacked by an OpenAI testing agent, closed models refused to help analyze the attack corpus, leaving only open models usable—“Why do we let them be both the athlete and the referee?” Regulators should filter training data rather than user behavior, while closed models can be cut off by policy at any time. “However you look at it, having no open source is the most dangerous thing about this era.”

🔗 Original source & video: E246｜What Is Distillation? How Silicon Valley Sees Chinese Open Models Closing in on the Frontier

王铁震的期中考判断：中国模型迭代速度已经上来；在后续中国模型持续超过 GLM-5.2 的前提下，超过 Fable 只是时间问题。 GLM-5.2 has “basically replaced Opus 4.8” in his live workflow: Fable handles task decomposition, while GLM-5.2 does the work. Wang sees no paradigm shift in Fable; given enough compute, data, researchers and time, it can be built. The market conversation relayed by Raymond: Elon Musk said Chinese models would reach Fable’s level in Q1 2027, and 唐杰 replied, “It won’t take that long.” Zhipu shares rose “20% or 30%” the next day, hitting a recent high.

100% RSI (recursive self-improvement) is considered extremely difficult to reach: like entropy increasing, a fully autonomous iterative loop will produce an ever larger mess of code until complexity exceeds the model’s context and intelligence. The system must continuously ingest low-entropy, high-value human input from the outside world. Raymond paraphrased an Anthropic paper as saying that 80% of the code in Claude Code is written by Claude Code itself; Wang estimates that “perhaps 60%” of that still reflects direct human instructions. One potential RSI metric is a gradual decline in the share of human participation, although the original wording later says the ratio will “forever approach 100%”; the direction of the ratio is ambiguous.

The model’s fundamental ceiling is that “the corpus a model consumes is, in practice, its shackle.” A model can transfer methods from field A to field B, but creating an entirely new field C with a completely independent method looks like a hallucination under its training objective. Humans can compress complex systems into invariants and modules through abstraction; that is why researcher taste is valuable and why Recursive’s large financing round makes sense. For now, RSI is at least easier to push along the efficiency frontier.

Wang characterized Fable’s national-security-style launch as “very good marketing,” not a paradigm-level leap. Using Kimi K2.5 and K2.6, he found the FreeBSD vulnerability promoted by Fable after being given a hint. His guess is that the gap lies mainly in offensive-security data and model size, not in the paradigm. “The kill line for open-source models has moved up,” and Fable now has to create meaningful separation and tell the story well. Raymond therefore compared Anthropic to having trained “an Einstein gone dark.”

Open source is commoditizing models, with a direct investment implication: “Moutai will still be No. 1 and command a premium; but it will be hard for the baijiu in second place to raise prices.” China has at least 10 companies capable of building models, including Meituan and Xiaomi. With too many players and free models, even the leader may not be able to keep raising prices over time. Open source served as a “100 into 10” filter in the primary market, but “investing in the eighth-place baijiu is a bit stupid.”

China has “kill lines everywhere,” while data silos could push the industry toward “one big model per capita.” Meituan trained the 1.6T-parameter LongCat 2.0 on a 50K domestic-chip cluster, and non-model companies are continuing to raise the floor. Because data across Chinese apps is not connected, Raymond worries that every major company will build its own model. Wang agrees that data silos are one reason each company needs its own model; the US has similar dynamics, with xAI’s acquisition of Cursor tied to its programming data.

Hunyuan 3 is the quarter’s key signal: its architecture is unchanged from the preview, yet swapping in new data and training methods delivered a huge performance jump. Wang’s reasonable outside-in inference is that architecture research is mainly about lowering costs, not necessarily directly improving capabilities. That in turn suggests data may matter enormously. Tencent has data channels through WorkBuddy, WeChat and Official Accounts; if it can use them appropriately and legally, Hunyuan 3.5 and Hunyuan 4 could still deliver qualitative leaps. Hunyuan 3 is smaller than GLM-5.2 and can run on a single domestic machine with 8 GPUs.

Raymond’s non-consensus view: “Coding” is the wrong word; it should be called “Automation.” Every model company is moving into coding because it covers work with relatively high, replaceable hourly value, underpins many digital tasks, and offers a natural feedback loop through compilation and execution results. The TAM is not programmer compensation but every automatable interaction in the digital world. On whether OpenAI should spend $60B to acquire Cursor with its own stock, Wang says Sam’s motivation is weaker than Musk’s, given Musk’s idle compute capacity; Raymond thinks the deal is mainly about data. The Q3 questions remain whether Anthropic and OpenAI will “very likely” go public, what their financials look like, whether AI stocks will be re-rated, and where the kill line ultimately settles.

🔗 Original source & video: 43. “26Q2” Large-Model Midyear Report: RSI, Fable, 5.2, Hunyuan 3, LongCat—Kill Lines Everywhere — 王铁震

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.”

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

Wang Tiezhen’s core view is not that open source must win: closed source sets open source’s capability ceiling, but the gap will not steadily widen. He agrees with the first half of 李彦宏’s argument: if a closed-source company has no product better than a free open-source model, its pricing model cannot last; but once both open- and closed-source models clear the GPT-4-level usability threshold, competition will shift across niches such as coding and writing, making it difficult for any single leaderboard to prove that one side has left the other “far behind.”

Llama 3’s significance is that it puts near-GPT-4 capability in the community’s hands, not that enterprises will immediately deploy the largest 400B model. The released 8B and 70B models and the then-unreleased 400B model sit on different cost curves: within the same family, with the same training method and architecture, larger models are necessarily more capable and more knowledgeable, but inference is more expensive and slower, and ordinary developers may not even be able to run 70B. Wang’s view is that the industry will “scale up first, then make models smaller”; mass deployment is more likely to rely on models below 10B, including 2B, 4B and 8B models on edge devices.

Open and closed source are not binary camps but a broad spectrum running from fully open training to open weights to closed-source services. A company can layer private data on an open-source model and charge for a closed-source product at the “last mile”; it can also release a smaller model for user acquisition while commercializing a more powerful version. Wang argues that sustainable open source usually needs a path to self-funding: “Only open-source projects that can ultimately convert into a business model can go further.”

Open source’s commercial value can come from distribution, ecosystems and complementary assets, and it can also support licensing or service revenue. Alibaba can use open source to drive cloud compute and GPU consumption, hardware companies can use software releases to expand device demand, and Meta may use model ecosystems to supplement content and interaction in the metaverse. Open source can also function as a market strategy, technology brand and recruiting tool, so “free” does not mean there is no business model.

Open source’s real contribution is to compress the cost of experimentation and collaboration: one person need not understand all the math, own a large GPU fleet or even work for a particular company to improve one part of the stack. Wang cites an analysis showing that large-model inference costs fell by roughly 100x over the past year, with much of the work driven by the open-source community. Projects can also be copied, dissected and developed into new projects, allowing A and B technical paths that a company would have to choose between to evolve in parallel in the community. “Everyone stands on someone else’s shoulders” (「每个人都站在别人的肩膀上」) is why he believes open science accelerates innovation.

Open weights are not the same as full open source, and Llama is closer to open access. Full open source would disclose the dataset, cleaning methods, training and inference code, model architecture, logs and weights; releasing only the weights is like providing an executable .exe without the source code. Developers can still fine-tune, merge models, quantize them or extend the context window, but they cannot fully reproduce how the model was trained.

The safest path for a founder is to validate demand with the strongest closed-source model first, then decide whether to migrate to open source. Wang recommends testing prompting with models such as Claude 3 Opus; if prompting solves the problem, there is no need to rush into fine-tuning, since prompt costs are far lower and a product may launch within a week. Once real queries accumulate, the team can organize the data for SFT and private deployment to reduce costs and protect privacy. Most startups may be able to afford SFT, while continuous pretraining can cost more than 10x as much and requires knowledge of the original developer’s training-data mix and other details.

Compute remains the hard constraint behind the open-source ecosystem, and distributed collaboration will not eliminate talent hubs. Llama 3 was trained, according to the episode, on 24,000 GPUs and 15T tokens; purely community-funded grants may reach only the hundreds-of-GPUs scale, with thousand-GPU support rare. At larger scales, fault replacement, checkpointing and automatic recovery become systems-engineering problems. Hugging Face can operate globally through remote work and asynchronous Slack collaboration, but Beijing, Silicon Valley and Paris still form hubs through talent density: “Talent hubs will always exist” (「人才高地一定会存在」).

🔗 Original source & video: 01. Taking on 李彦宏 as Llama 3 Launches: What Is the Open-vs.-Closed-Source Debate in Foundation Models?