Moonshot's Kimi K3 license splits open weights from free commercial use, and most of China's open-weight labs have not gated theirs

Key takeaways

For two years, most analysis has treated Chinese open-weight models as a single force pushing in one direction: dragging the price of AI toward the floor. The strongest version of that view, and one SecondSource has tracked for a long time, comes from the investor Bill Gurley. In his account, Chinese policy makes a monopoly at the model layer impossible, margins there are structurally crushed, and every Chinese open-weight model shipped abroad puts a ceiling on what frontier labs everywhere can charge. Open weights have never automatically meant free commercial use; the license decides that. On July 16, Moonshot AI, the Chinese lab behind the Kimi models, released Kimi K3. On July 27 the weights went up as promised, and the license and price list published alongside them put a condition on that view. SecondSource checked the license texts and the price lists. Three conclusions follow:

• Kimi K3 formally separates "open" from "free for commercial use." Under the K3 license, any company that runs a model-inference service and whose group revenue exceeds US$20 million a year must sign a separate agreement with Moonshot before commercial use. ChinAI Newsletter, Jeff Ding's newsletter translating Chinese-language writing on AI, computes a blended price per million tokens for typical usage. On that measure K3 costs US$2.30, the very top of the Chinese price band and about 13 times DeepSeek V4 Pro's US$0.18. Both companies' official price lists confirm that ratio. The open weights spread the model, and the license and the price list decide what Moonshot collects for it. No Chinese frontier lab had paired a revenue-gated license with a top-of-band price before. MiniMax gated earlier, but it prices M3 at US$0.22.

• The expected industry-wide turn did not happen. SecondSource compared the license text of K3 and nine other Chinese labs word for word. Besides Moonshot, only MiniMax, the lab behind the M-series models, gates commercial use, and MiniMax did so more than a month before K3 and went further. DeepSeek and Zhipu remain unconditionally open. Zhipu, the GLM lab also known as Z.ai, puts "Pure Open: An MIT open-source license — no regional limits" on the first page of its model card as a selling point. Under one label, three different business designs have diverged. Moonshot and MiniMax put the gate in the license, Alibaba and OpenAI's gpt-oss put it in what gets released, and Zhipu and DeepSeek stay fully open and make money elsewhere.

• Gurley's price-crush argument has not been overturned, but it now holds only while at least one capable Chinese lab stays on a permissive license at a low price. As long as Zhipu and DeepSeek stay there, the pressure stays. K3 is simply the first model trying to climb out of the low-price band. Whether it can is the biggest unknown: two days after launch Moonshot suspended Kimi subscriptions for lack of compute, and the public evidence cannot yet tell whether the high price is a value proposition or rationing by another name.

What the license actually says

On July 27, the K3 license, not the benchmark scores, dominated the discussion in English-language AI research circles. Nathan Lambert, who writes the Interconnects newsletter, posted a clause-by-clause breakdown that drew more than two hundred reposts that day. The reason for the attention was that the document rewrote an industry convention.

SecondSource pulled the full license text from two independent copies, on Hugging Face and on GitHub, and compared them word for word; the two copies match apart from one whitespace character (Moonshot AI). Several details that had rarely been reported in full are worth laying out.

The first gate targets companies that sell inference. Any company operating "model as a service" must sign a separate agreement with Moonshot before commercial use, once the company and its affiliates together pass US$20 million in revenue in any rolling twelve-month period. Three choices of wording matter. The test counts aggregate group revenue, not revenue from the model-serving line of business. It is measured on a rolling twelve months, not once on the release date. And once the threshold is crossed, what stops is commercial use of any kind, not just resale of inference. The license's definition of "model as a service" also deliberately leaves out two kinds of business: end-user products that wrap the model's capability inside a feature, and intermediaries that simply forward requests to a model hosted elsewhere. The gate is aimed squarely at cloud providers and inference vendors. Application developers are outside its range.

The second gate is an attribution duty. A commercial product with more than 100 million monthly active users, or more than US$20 million in monthly revenue, must prominently display "Kimi K3" in its interface.

There is also an exemption that rarely comes up in the English-language technical discussion: use of K3 through Moonshot's official products or through "certified inference partners" is exempt from both gates. That writes a channel-certification scheme directly into the license. An inference vendor that wants to sell K3 legally either negotiates with Moonshot itself or joins Moonshot's certified network.

Just as important is what the document does not contain: no geographic restriction, no use restriction, no termination clause, no retroactive clause. A year earlier, Tencent's Hunyuan license explicitly barred use in the EU, the UK and South Korea (Interconnects). The K3 license is not defensive in that way. It is built to monetize: apart from requiring large players to come and sign, it regulates nothing.

The K3 license is also a newly drafted document, not a patched old one. The license for the previous model, Kimi K2, was titled "Modified MIT License," and its only change to the standard MIT text was the attribution duty. The K3 license is titled "Kimi K3 License," and a word-for-word comparison shows it does not keep MIT's grant-and-disclaimer structure; the text was rewritten. From K2 to K3, Moonshot moved from asking large users for credit to requiring a signed agreement from any licensee big enough to cross the line. Moonshot's K2 and K3 licenses are the cleanest before-and-after pair so far of one lab moving from a permissive license to a gated one.

The price list, published the same week, is the other half. ChinAI Newsletter, Jeff Ding's newsletter translating Chinese-language writing on AI, computes a blended rate for typical usage: 70 percent of tokens are cached input, 20 percent fresh input, 10 percent output. On that basis K3 costs US$2.30 per million tokens. On the same ruler, Alibaba's Qwen3.7 Max costs 1.40, Zhipu's GLM-5.2 0.90, MiniMax's M3 0.22 and DeepSeek's V4 Pro 0.18 (ChinAI Newsletter). K3 is about 13 times DeepSeek. On output price alone, K3 is up three and a half times from Moonshot's previous model, K2.6. That multiple and the K2.6 baseline both rest on a single relayed account.

ChinAI also translated an article from the Chinese industry outlet Diyi Xinsheng, which describes the pricing bluntly: "By matching the performance of flagship overseas models while anchoring prices to the ceiling for domestic models, [K3] drives China's large models to move beyond 'cutthroat price competition' toward 'value monetization.'" "Domestic" here means models made in China.

SecondSource checked those numbers. The official price pages of Moonshot and DeepSeek both list the original yuan prices, and converted to dollars they land within 10 percent of the figures above. The key 13-times multiple therefore stands on two official price lists rather than on a relayed account. The official prices for Qwen, Zhipu and MiniMax sit behind login walls and have not been checked directly. Zhipu's figure matches the order of magnitude in an independent account by another US analyst; the other two rest on the single relayed account.

Why the gate and the high price are one design

Read either document alone and Moonshot's move looks different from what it is. Put the license and the price list side by side, and they turn out to be two parts of one mechanism.

Start with what would happen if Moonshot raised prices without a gate. The cost structure of serving an open-weight model is public. By an estimate SecondSource recorded earlier, self-hosting a model of K3's class, around 2.8 trillion parameters, takes a rack of the GB200 NVL72 class. Buying and installing one costs about US$3 million to US$4 million, and renting one runs about US$7 million a year (Exponential View). That bill is astronomical for a startup and a rounding error for a cloud provider, so only large players self-host. The figures are one analyst's engineering estimate. The rental figure usually includes power and operations, so the two figures sit on different cost bases. Once the official API charges a high price, the customers most able to pay will download the weights and host them, or hire a cheap inference vendor to host them, and the official API's revenue leaks away.

The reverse fails too. A gate without a higher price holds the negotiating table but leaves nothing worth selling at it. Only the two together work. Openness spreads the model; the license pulls the large players it cannot stop back to the negotiating table. Only then can the API's high price hold, because self-hosting is expensive and, once the rack is up, the self-hoster still has to sign.

The legal form of "open weights plus a commercial gate" is not Moonshot's invention. Since Llama 3, Meta's Llama license has carried a clause requiring any company with more than 700 million monthly active users in the month of release to request a license from Meta, which Meta can grant or refuse at its sole discretion. The two thresholds serve different purposes, though. Llama's 700 million threshold is a one-time snapshot on release day that only four or five giants in the world can hit; its point is to stop rivals on the scale of Google or ByteDance from taking the model for free. K3's US$20 million threshold is tested on a rolling basis, and nearly every commercial inference vendor will hit it. Both are called gates, but they do different jobs: Llama's keeps a handful of giants from free-riding, and K3's collects from the whole inference-supply layer.

What Moonshot's leadership says supports the reading that this is design, not oversight. In a recent interview, Moonshot's founder Yang Zhilin was asked by the interviewer, Zhang Xiaojun, whether Moonshot would stay open source for the long term. Yang answered: "That's what we hope to do for the long term, but it doesn't have to be open source only." He added: "For example, in collaborations with certain companies, not everything will be opened up." The interview reached English through ChinAI's translation; the Chinese original hasn't been checked directly (ChinAI Newsletter). The license gate is that sentence turned into law.

The peer prices, both Yang quotes in this piece and the subscription suspension all come through ChinAI, a single English-translation channel, and are the most concentrated sourcing in the piece. The license text and the official K3 and DeepSeek price lists are primary documents SecondSource retrieved directly. The two kinds of evidence do not carry the same weight.

The other nine licenses: the expected industry-wide turn did not happen

If K3 marked a shift in Chinese open weights from free distribution to license monetization, other labs should show signs of following. SecondSource pulled the license text for the strongest current open-weight model from each of ten Chinese labs and compared them one by one. The results are below. Prices use the blended rate above. The capability column is Artificial Analysis's composite Intelligence Index; Artificial Analysis is an independent firm that benchmarks models and publishes that index. All readings are from July. Cells resting on a single relayed source are marked with an asterisk.

The table shows three things.

First, the worry that K3 is a one-off is settled: there is a second case, and it went further. MiniMax, the Chinese lab behind the M-series models, changed its license far more sharply than Moonshot. Its M2, in December 2025, was still under a modified MIT license that only asked very large products for attribution. M2.7, in April 2026, switched to a non-commercial license: any commercial use, by a company of any size, needed prior written authorization, with no revenue threshold as a cushion. M3, in June, stepped back to "attribution plus notice for commercial use, and prior written authorization above US$20 million in yearly revenue." After the tightening and the loosening, the dollar line landed almost exactly where K3's sits: US$20 million in annual revenue. The reach is different, though. K3's gate hits only inference-service operators and explicitly exempts the application layer; M3's gate applies to any commercial use. The two match on the dollar line, not on scope. As for who influenced whom: the M3 license (June) was public more than a month before the K3 license (July 27). Public evidence cannot show who took a cue from whom, but it does establish that MiniMax did not copy K3.

Second, the more important signal points the other way: the main suppliers did not move, and they use not moving as a weapon. DeepSeek's V4 Pro is plain MIT, with no attribution duty at all. Zhipu's GLM-5.2, from the lab also known as Z.ai, is plain MIT too, and the model card puts that in the first row of its key features: "Pure Open: An MIT open-source license — no regional limits, technical access without borders." With peers starting to gate, that line is competitive language with a clear target.

Alibaba takes a third route. Its flagship, Qwen3.7 Max, shows no weight release: as of publication, Alibaba's official Hugging Face account has no matching model, and Alibaba has made no open-weight announcement. The flagship is API-only, and Alibaba's open weights stop at the second-tier Qwen3.5-397B, under plain Apache. One claim circulating in the market needs correcting here. On July 19, Alibaba's official Qwen account did pledge that the 2.4-trillion-parameter Qwen3.8 was going open-weight soon. As of publication, though, Hugging Face, GitHub and ModelScope, Alibaba's own model-hosting site, show no weights and no license terms. A pledge and a release belong in separate records; for now the pledge counts as evidence for neither side.

So "Chinese labs are turning, as a group, to license monetization" does not survive the evidence. Of ten labs, two gate commercial use. The three biggest suppliers, DeepSeek, Zhipu and Alibaba, put no gate in their licenses at all.

Third, step back: under the label "open weights," three business designs have diverged.

The first design puts the gate in the license. The weights are fully released, and the license terms pull large players back to the negotiating table. Moonshot and MiniMax both do this. There is precedent outside China. Black Forest Labs' image model Flux splits its license into three tiers: the smallest version fully open, the developer version non-commercial, the flagship by negotiation. Stability AI's Stable Audio runs a community track and an enterprise track, with an enterprise license required above US$1 million in annual revenue (Stability AI). Both have the same "small is free, big pays" shape.

The second design puts the gate in what gets released. The license itself is clean, but the flagship's weights are withheld, or only a cut-down version ships. Alibaba withholding its flagship is one case. OpenAI's gpt-oss is another: it uses an unconditional Apache 2.0 license but deliberately does not release the base model (Interconnects / Substack). The base model is the raw version before chat tuning, and only with it can others retrain freely.

The third design is genuinely fully open, with the money made elsewhere. Close to three-quarters of Zhipu's 2025 revenue came from deploying models in customers' own data centers (Epoch AI). For Zhipu, open weights are the front door for enterprise customers, not a leak that lets others take the product for free. DeepSeek is backed by a quantitative hedge fund, and its rock-bottom price is itself the strategy. The three designs have entirely different business consequences, yet under the label "open weights" they are indistinguishable.

If the reader keeps one call from this piece, it should be this: as a unit for business analysis or policy classification, the label "open weights" has stopped working. A joint letter launched on July 24 and since signed by more than 150 US organizations, Google, Meta and OpenAI among them, still uses "open weights" as the unit of its policy ask (Microsoft). Under that one term, what the signatories want protected runs from unconditional Apache licenses to custom terms with a toll booth.

Why, then, did these two labs gate? Look at where their money comes from. In April, the industry commentator Nathan Lambert singled out Moonshot, MiniMax and Zhipu as labs that, on their current strategies, would come under financial strain over the next few years, because training costs are moving from the millions of dollars toward the billions (Interconnects).

Four months later, all three labs Lambert named have moved, in different directions. Zhipu found other money through its stock listing and its deployment business, and kept its license fully open. Moonshot and MiniMax both rest their revenue on the model layer itself, and their licenses grew gates. Our read: whether a lab gates depends less on where its capability stands than on how it makes money. A lab needs a license to protect its model API only when selling that API is its main income. Capability position decides how high a price can be stacked on top of the gate. That brings the argument back to K3's 13-times price.

What Gurley's price-crush argument gets wrong, and what it gets right

Now set the K3 case against the view SecondSource already holds. Gurley's argument is a complete causal chain. China's five-year plan explicitly promotes open source, and the provinces compete with one another, so several models always coexist on the market. Models also differ from electric vehicles: a carmaker cannot use a rival's car to improve its own, but one model can be distilled from another, so convergence is faster and margins are harder to defend. Distillation means training a model on large volumes of a rival model's output, at a fraction of the cost of original training. Gurley's conclusion: the model layer is permanently commoditized, value stays in the distribution layer, and Chinese open weights shipped abroad put an invisible ceiling on what frontier labs everywhere can charge. The source is an interview Gurley gave on the BG2 podcast after a visit to China in August 2025 (BG2 Pod).

That chain has a record that is hard to deny. China's price war is real history. DeepSeek set it off in May 2024; ByteDance, Tencent, Baidu and Alibaba followed, some of them cutting below cost (Hugging Face Blog). Today DeepSeek V4 Pro's official output price is less than 4 percent of the list output price of Western flagship models, according to DeepSeek's July price list (DeepSeek).

The ceiling pressure also has a named observation. The investor Jamin Ball recorded that after Zhipu's GLM-5.2 rose on frontier capability at a much lower price, OpenAI priced its new model Luna to match it (X / @jaminball). The source is Ball's own post; no permanent link to it is available, and it is a single-source observation. If it holds, it is the first named case in 2026 of Chinese open weights setting a price and a Western closed model following.

So what did K3 overturn? This piece's call: K3 leaves the mechanism standing but narrows where it comes from. The ceiling pressure was never applied evenly by "every open-weight lab." It was applied by the few near-frontier labs still on permissive licenses in the low-price band. As long as Zhipu and DeepSeek stay fully open and cheap, Western flagships feel exactly the same pressure. K3's climb out of the price band changes Moonshot's own income statement, not the market's price structure. Gurley's chain is still alive at the market level, but it has a condition nobody had written down before: the mechanism runs only if at least one sufficiently capable lab stays in the permissive, low-price band. That condition has moved from "taken for granted" to "needs watching," and that is the most substantive correction the K3 case makes to SecondSource's existing view.

Yang Zhilin has said as much. Zhang Xiaojun reminded Yang that a year earlier Yang had said, "The leader won't open-source; only the laggards will," and asked why Moonshot had now open-sourced. Yang's answer: "Because right now, globally, we're not fully in the lead yet (laughs)." The exchange comes from the same interview, in ChinAI's translation (ChinAI Newsletter). By that logic, every lab holding up the mechanism in the low-price band today is keeping its own exit route open.

Another passage from the same interview is often cited against Gurley's chain, but it does not reach it. Yang said plainly that community feedback cannot improve the base model. On making the model stronger, his words were: "currently only the original developer can do that." The value of open source, in Yang's telling, is "more about empowering downstream applications than feeding back into the improvement of the base model." That rules out a popular idea, that an open-source community will jointly push the underlying model forward. Gurley's argument is about something else: one-way distillation between rival labs. That requires no community collaboration at all, only that a rival's model be available.

SecondSource previously recorded an estimate that points one way. Training a model up to the same level with reinforcement learning on its own costs about 17 times the compute of distilling a rival's finished model (Interconnects / Substack). Followers who take the shortcut save a great deal. The distillation leg of the price-crush argument still stands; what Yang cut away is only the more optimistic version in which everyone pulls together.

What the evidence cannot separate: premium price or rationing

Everything so far concerns what Moonshot intends. Whether Moonshot can pull it off hangs on one fact. K3 launched on July 16. Two days later, around July 18, new users flooded in, and Moonshot suspended Kimi subscriptions and capped usage for existing users. Two outlets reported this independently. The Chinese industry article relayed by ChinAI attributes it to a compute shortage; Bloomberg, as relayed by Stratechery, frames it as demand overload. The two accounts are compatible, because demand overload on fixed capacity is a compute shortage (ChinAI Newsletter; Stratechery). The limit to keep in mind: what was suspended was consumer subscriptions, and no public evidence shows the state of capacity on the API side.

That gives K3's 13-times price two opposite readings. The first: the price is a value proposition, and the capability earns it. Artificial Analysis's July Intelligence Index gives K3 a 57, at the tail of the closed flagship group and at the head of open-weight models. On the front-end web generation sub-leaderboard of Arena, a platform that ranks models by blind human head-to-head votes, K3 ranks first as of mid-July. Both data points come from a roundup by Latent Space, the AI-engineering newsletter; no permanent link to the leaderboard pages is available (Latent Space). On this reading the compute shortage is a growing pain.

The second reading: the price is a rationing tool. Serving capacity is simply not enough, and the high price exists to squeeze demand down to the compute available. Once compute frees up, the price comes down. Moonshot has reportedly told investors it plans to list in Hong Kong within six months to raise money, according to Don't Worry About the Vase, Zvi Mowshowitz's newsletter, citing Bloomberg (Don't Worry About the Vase). The two readings carry opposite implications for Moonshot's valuation and for whether "value monetization" is a real trend, and the current public evidence cannot separate them.

SecondSource has followed the same structure on the Western side for half a year. After model makers' token sales turned gross-margin positive in the first half of 2026, the argument has been over what those new margins are. Are they real pricing power, built on a compute shortage plus a widening quality gap between models? A temporary rent collected during a shortage? Or is the demand signal itself inflated? This publication's July 20 deep dive reached a layered conclusion: pricing power exists only when three conditions hold at once, namely the very top flagship tier, a continuing compute shortage, and buyers who have not yet worked out whether their AI investments pay back (verdict date July 20, 2027). K3 carries the same question, unchanged, into China. Its top-of-band price was likewise born in a shortage, and likewise has not been tested in a period of ample compute.

The tree in one sentence: once model makers' token sales turned positive, three positions face off over whether the margins last, and K3 is the first case on the Chinese open-weight side to take a stand on this tree. K3 is aiming for the pricing-power branch, but the current evidence cannot rule out the reading that its price is only shortage-era rent.

One thing, though, is not clouded by this confusion, and it happens to be the hardest evidence in the K3 case: the license. The license terms do not depend on compute supply, so they stay evidence even if the price falls. Even if K3's price is cut in half next year, the "separate agreement above US$20 million in revenue" structure stays in place, and so does the certified-inference-partner channel. This piece treats that as the dividing line for evidence strength across the whole case: "Moonshot wants to take back control of monetization" is hard evidence, at the level of a primary legal document; "K3 can hold a top-of-band price" is still undecided.

Our call

The label "open weights" can no longer be used to predict business outcomes. Under that one label there are now three designs: the gate placed in the license (Moonshot, MiniMax), the gate placed in what gets released (Alibaba, OpenAI's gpt-oss), and fully open with the money made elsewhere (Zhipu, DeepSeek). What actually predicts outcomes is where the license gate sits and how the lab makes money. The price-crush mechanism of China's model layer still holds at the market level, but its condition must now be written out: ceiling pressure exists only while at least one sufficiently capable lab stays in the permissive, low-price band. K3 is the first model to climb out; Zhipu and DeepSeek are the labs holding the mechanism up today. On K3 itself, the license is settled. The price was set during a compute shortage and has not been tested since. Public evidence cannot yet tell whether it is value pricing or a price used to ration scarce capacity.

What would prove this wrong (the observation windows are set by SecondSource unless noted, and checked when they expire). One: before mid-2027, K3's actual transaction price falls back into DeepSeek's band (a blended price below US$0.50), or Moonshot cuts prices or loosens the license gate because of weak adoption. That would mean the "value monetization" reading fails and the rationing reading wins, and this publication's core call here would have to be revised down. Two: after Moonshot lists and raises money and expands compute (with resumed subscriptions as the signal), K3's price still holds at the top of the band. The rationing reading weakens, and the value-proposition reading gains. Three: the next flagship open-weight release from Zhipu or DeepSeek also grows a revenue gate. The condition that someone is still holding the permissive low-price band is shaken, and the market-level conclusion about the price-crush mechanism needs re-examining; conversely, as long as both stay plain MIT, K3 remains an exception. Four: public evidence that any cloud provider has signed a separate agreement with Moonshot or joined its certified inference partners. That would be the direct signal that the license gate has moved from paper to enforcement. No public channel exists to check this today; if Moonshot lists, its prospectus is the most likely place for it to appear. Five: Alibaba has publicly pledged that the 2.4-trillion-parameter Qwen3.8 is going open-weight soon but has not released it. When it does ship, whether it carries Apache or a custom gated license is the cleanest single test of "is gating an outlier or a turn," and it is also the most likely of the five to come due first.

What this changes, and for whom

Product and application teams. The good news comes first: end-user applications that wrap the model's capability inside product features are explicitly exempt under the K3 license and can ship as usual. Three architectures, though, can turn an application company into a "model as a service" operator without anyone noticing: a forwarding endpoint that lets customers pass their own input parameters, prepaid token bundles, and white-labeled API resale. That makes the license an architecture decision, not something legal can patch afterwards. Because the threshold counts group revenue over a rolling twelve months, a license check made before launch costs far less than one made after revenue has grown to the line. For products built on front-end generation, K3 has named evidence of leading in that use. For other uses, K3 has no named evidence of a lead yet, so a price gap of 13 times has no business case.

Enterprise procurement and platform engineering. Evaluations of "self-host an open model to cut costs" gain a new first step: reading the license, and reading it for the company's own position. An application company (usually exempt) and an inference reseller (almost certainly over the line) are in completely different situations under the same license.

Investors. Three questions now need separate answers. First, the case that Chinese open weights will drive model-layer prices to the floor still holds, but only while Zhipu and DeepSeek remain in the permissive low-price band. Second, in Moonshot's Hong Kong listing story, the "monetization" half still rests on soft evidence: whether the price holds after the listing funds more compute is the moment of truth. Third, whether a lab gates its license is a telling signal of monetization pressure. MiniMax tightened and then loosened its license within half a year. That path shows the pressure more plainly than a company presentation would.

Policy circles. "Open weights" as a unit for regulation or exemption is no longer fine-grained enough. Under one label sit unconditional Apache licenses and custom terms with a toll booth; a line drawn along that label misses the variables that actually matter (where the gate sits, what gets released, where the money comes from). The license research also turned up two unconfirmed leads. An investment bank's research report is said to suggest that Chinese AI labs may move toward a "paid weights" model, and China's Ministry of Commerce is reported to be weighing limits on downloading model weights from abroad. Neither is confirmed, and this publication does not cite them as fact for now; SecondSource plans to check the primary documents and report back within August 2026. If they hold, license gating could be an early sample of an industry-level turn rather than two labs' separate business choices.

US frontier labs. K3 climbing out of the low-price band is a local positive for US flagship pricing, with one fewer source of downward pressure. But the main ceiling pressure has always come from Zhipu and DeepSeek, and they have not moved. The thing to watch is the license on Qwen3.8, the 2.4-trillion-parameter model Alibaba has pledged to release as open weights. The license that the next frontier-class Chinese open-weight model carries will decide whether the ceiling keeps pressing down or loosens across the board.

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.

Sources cited in other language editions

These sources are linked in the body of this piece's other language editions but not in this one.

• huggingface.co/MiniMaxAI/MiniMax-M3/commits/main