⚠️ Editorial note: The open source ecosystem in China operates under a distinct institutional framework — characterized by state-led initiatives, intranet-like boundaries, and top-down governance. Readers should be aware that this context differs from the community-driven open source model common in other regions. The term “open source” as used in Chinese media may refer to practices that diverge from the conventional definition.

China Open Source Daily — 2026-10-23

🏛️ Policy & Regulation — Six Characters That Make Publishing the Cheapest Licence: Reading MOFCOM Announcement No. 48 in the Original Chinese

1. The instrument, and the date that matters. On 2026-10-16, China’s Ministry of Commerce (MOFCOM) and Ministry of Science and Technology (MOST) jointly issued Announcement No. 48 of 2026 (商务部 科技部公告2026年第48号), formally titled “Soliciting public comment on the revision of the Catalogue of Technologies Prohibited or Restricted from Export, and implementing interim registration management for the export of relevant technologies.” The announcement opens by citing two statutes — the Foreign Trade Law and the Regulations on the Administration of Technology Import and Export — and then does three things at once.

First, it proposes 40 additions across three fields. Of these, 23 are AI technologies: large models (大模型), agents (智能体), compute clusters (算力集群) and information-processing technologies. Five are biomedicine — cell and gene therapy, small-nucleic-acid drugs, pathogenic microorganism technologies. Twelve are photovoltaics — polysilicon manufacturing, wafer manufacturing and cell manufacturing, with the scope explicitly extended to “the key and core equipment that carries the relevant technology.” The comment period runs 2026-10-16 to 2026-11-15.

Second — and this is the procedural novelty — it switches on an interim pre-export registration regime with effect from the day of publication. The text is unambiguous: “from the date of publication of this announcement until the date the revised Catalogue is formally revised and implemented” (自本公告发布之日起，至《中国禁止出口限制出口技术目录》正式修订并实施之日止), any technology-export operator must register with MOFCOM’s unified business-system platform before exporting any technology listed in the annexes. Exports already begun but not yet completed (已经开始但尚未完成) fall inside the registration scope. Four registration triggers are specified: transfer of technology ownership abroad (e.g. an IP sale); licensing domestic technology for use abroad; a change in the owner or actual controller of domestic technology to a foreign entity; and any other circumstance MOFCOM later designates by notice.

Third, it carves out exemptions — and this is where the open-source question lives. Two exemption clauses appear in the announcement’s first section, and they are not identical. For biomedicine: “for outbound licensing models that do not involve transfer of ownership, facilitation measures such as automatic licensing (自动许可) apply.” For AI: “technologies that have entered the public domain by means of open sourcing or the publication of academic papers (通过开源、发表学术论文等方式) are not included in the scope of management (不纳入管理范围).”

2. The distinction between “not in scope” and “automatic licence” is the whole finding. These are two different legal animals, and the difference is not cosmetic. An “automatic licence” is a licence: a permission that is granted with reasonable speed. “Not in scope” is a scope exclusion: the technology is not subject to the regime at all. MOFCOM’s own instrument therefore designates, for the category that matters most commercially, the stronger of the two reliefs — and it grants that relief to the one act that no regulator can reverse.

There is also a second reference to open sourcing, reported by ChinaTalk’s three-day reading of the full draft and its attached Q&A, which governs technology already listed as restricted: for any new or unreleased technology within the AI section, open sourcing requires an automatic licence, described in the Q&A as “basically achieving immediate report, immediate release” (基本实现即报即出). Read together with the scope exclusion, the structure is: already public → outside the regime; not yet public, restricted category → inside the regime but on the fast track; not yet public, everything else → ordinary licensing. A third reader-facing consequence is documented: the exemption is what gives firms a timing incentive — anything a company publishes before the revised Catalogue takes effect is permanently outside it, whereas a release after that date travels through a filing desk.

3. What the coverage does and does not reach. Per The Next Web’s reading and ChinaTalk’s, the curbs cover weights, source code and other core components for models meeting any of three thresholds — ≥1e26 FLOPs, ≥2 trillion parameters, or “capability at the global leading level” — and also extend to AI chip design know-how at 14nm or below or using 2.5D/3D stacking, and to certain memory designs. Serving foreign users through apps, websites and APIs remains permitted; running a covered model on servers abroad counts as an export, though firms operating their own overseas cloud may receive an automatic licence. The MOFCOM spokesperson’s framing, quoted by TNW: the rules aim to “prevent misuse and abuse” and “ensure that AI remains under human control,” and do not target any country or firm.

Institutional Analysis & Commentary:

The published reaction to Announcement 48 has concentrated on the 40 items. That is the wrong object of attention, and the announcement’s own construction says so. The 40 items are a comment-period draft. The registration regime is already in force. One of these two facts has a legal effect today, and it is not the one receiving coverage.

Begin with the design of the registration trigger, because it is unusually well-engineered. Three of the four triggers — ownership transfer, outbound licensing, and change of controlling owner to a foreign entity — share a common property: they all regulate the movement of rights rather than the movement of goods. A conventional export control on a physical input is enforceable at a border. A model weight is not crossable; it is downloadable, and once downloaded it is everywhere. So the instrument does not attempt to police the download. It polices the transactional events through which a Chinese entity’s rights over the technology leave Chinese jurisdiction — and, critically, it treats redomiciliation as one of those events. The third trigger effectively codifies the lesson of a case the announcement does not name: a Chinese AI firm that relocates abroad no longer escapes, because transferring control of domestic technology to a foreign controller is itself the registrable act.

Now consider what an interim registration system with zero implementation delay is for. It is not an approval. Nothing in the announcement says a registration can be refused. It requires the exporter to disclose, before the fact, the operator and recipient identities, the destination country, the technology and its intended use, and the expected value. That is a census, and its value is informational before it is prohibitive. Stated in the language of institutional economics: this is a mechanism for converting an unobservable — what technology is leaving, and through whose hands — into a register, at near-zero administrative cost, ahead of the decision about what to actually ban. Hayek’s case for prices was that they aggregate dispersed information; the mirror case, and one that appears wherever states face fast-moving private knowledge, is that a filing requirement aggregates the same information without needing to price it. Announcement 48 builds the register first and decides the rules second, and the sequencing is deliberate.

Which brings the analysis to the exemption, and to why it deserves to be read as the centre of the document rather than a footnote. Open sourcing has been designated a lawful exit from an export-control regime. Not a mitigation, not an expedited review — a scope exclusion. Any firm holding a frontier model can, by publishing weights and code, place that technology permanently outside the Catalogue’s reach. And it must be permanently outside: a published weight cannot be recalled, which is precisely why the exemption was written for the past tense — “technologies that have already entered the public domain.”

The consequence is a price. Consider the choice facing a Chinese lab with a model above the thresholds, in the window between 2026-10-16 and the Catalogue’s formal revision. Option one: keep it closed, license it abroad, and accept a registration entry, a licensing desk, and a rule that may tighten. Option two: publish the weights under a permissive licence and the technology becomes unreachable by the control — permanently, and for every competitor on earth as well. The two options differ not in degree of compliance but in the identity of the party who ends up holding the asset. This is not a loophole accidentally left open. An instrument written by two ministries that had already read the alternative — the Chinese academic literature advocating a “negative list” for model releases — could have closed it with one sentence. It did the opposite, and the choice has an economic logic that is worth stating without either admiration or suspicion: if the state’s interest is diffusion, then a mechanism under which publishing is cheaper than licensing produces diffusion faster than any subsidy. The exemption is not a concession to open source. It is open source being used as a policy instrument.

The honest counterweight is the one a reader should hold alongside that reading. The exemption’s benefits are conditional on a licence being permissive, and on the licence actually being honoured. Publishing weights under a restricted-use or commercial licence — which is what Moonshot did with Kimi K3 — is not the same act as publishing under MIT or Apache 2.0, and the draft does not define “open sourcing,” so whether a restricted-use weight release qualifies for the exclusion is an open question. If the phrase is interpreted narrowly, the exemption rewards exactly one strategy: unconditional release. That is a consequential piece of unlegislated industrial policy, and it will be decided by whoever drafts the implementing guidance.

A third observation, offered as description rather than judgement. A control regime that exempts the public domain is structurally different from one that exempts a jurisdiction. It does not partition the world into permitted and forbidden recipients; it partitions technologies into published and unpublished. For a country whose stated objective in this field is to have its models adopted globally — a stated objective, not an inferred one; the State Council’s “AI+” action plan calls for helping Global South countries strengthen AI capacity — a published–unpublished boundary is the boundary that maximises reach while retaining leverage over the unpublished frontier. The regime’s selectivity is not about who may use the technology. It is about what counts as having given it away.

The checkable question. Three observables will settle what Announcement 48 actually is, and all three are public within a year. First: does the exemption survive contact with the implementing rules? If “open sourcing” is defined to include restricted-use weight releases, then open weights remain the default exit and the regime governs only the closed frontier. If it is narrowed to permissive licences, or to weights-plus-training-code, then the exclusion is far smaller than the announcement’s phrasing implies, and the phrase “not included in the scope of management” will have been a drafting convenience. Second: how many interim registrations are filed, and are any refused? A registration regime that has never refused anything is a census with an option attached; one that refuses is a licensing regime that has not yet published its criteria. Third: what happens to the models released inside the window? Every frontier model published between 2026-10-16 and the date of formal revision is outside the Catalogue forever. If that window turns out to have produced an unusual cluster of headline open-weight releases, the announcement will have functioned — whatever its drafters intended — as a publication deadline, and the correct name for the result is not a restriction but a deadline that made disclosure the dominant strategy.

Sources:

⚖️ Legal & Licensing — Apache 2.0 as Institutional Infrastructure: DeepInfer, and the First Day That Listed Huawei Beside Nvidia

1. The release. On 2026-10-20, DeepSeek open-sourced DeepInfer — the inference engine it uses to serve its own models in production — under the Apache License 2.0, at github.com/deepseek-ai/DeepInfer. Announcing it on X, the company framed the motive in cost terms: “We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.” The same day it published an updated deepseek-recipe, a companion library (Rust and Python) that converts API requests into model inputs and outputs into responses, so that self-hosted deployments can reproduce the behaviour of DeepSeek’s official API.

The technical inventory matters to the legal argument because it identifies exactly what has been transferred. DeepInfer bundles the company’s previously open-sourced kernels and communication libraries — FlashMLA, DeepGEMM, DeepEP — together with multi-chip inference distribution, speculative decoding, asynchronous scheduling and structured outputs. It is not a complete production stack: the public build supports offline inference, and per its own documentation several components needed for a full service are absent — a KV cache management system, cross-machine stage separation, and replica load balancing.

2. What changed since 2025, and why the licence matters more than the code. In April 2025 DeepSeek published a document explaining that it had considered releasing the entire engine and decided against it, on three stated grounds: the software had diverged too far from upstream vLLM to be maintainable by a small research team; it was “tightly coupled with our internal infrastructure, including cluster management tools, making it impractical for public deployment without significant modifications”; and maintaining another major open-source project was a burden the team could not carry. The 2026 release does not explain how those obstacles were overcome — but it does show the shape of the compromise: release the engine, withhold the harness. The bottlenecks that make inference an operational business (cache management, stage separation, load balancing) stay in house. What the community receives is the part that determines throughput on a given chip.

3. The hardware list, which is the institutional fact. Day-one supported platforms are Nvidia’s SM100 architecture (Blackwell datacenter GPUs) and Huawei’s Ascend 950 NPU series — and nothing else. Older Nvidia architectures, including Hopper, are conspicuously absent. Feature parity is not claimed either: batch-invariant inference and deterministic sampling are available only on Nvidia, and the published performance benchmarks use Nvidia hardware (four SM100 GPUs for prefill, 64 for decoding), leaving the Ascend comparison unstated.

The surrounding context is documented and recent. In September 2026, Bloomberg reported DeepSeek planned to deploy at least 160,000 Huawei Ascend 950DT chips at a new Inner Mongolia datacenter built specifically to run its models; The Information reported plans to increase Huawei use for training, with deliveries from Q4; Reuters reported DeepSeek and Huawei collaborating on software tools to make Ascend easier to program (2026-09-30).

Institutional Analysis & Commentary:

The interesting question about DeepInfer is not whether it is a generous release. It plainly is, and the motive is stated. The interesting question is what kind of asset an Apache 2.0 licence turns a production inference engine into once a Chinese lab is operating under an export-control regime that exempts the public domain.

Return to the previous section’s finding. Announcement 48 excludes from management scope technology that has “already entered the public domain by means of open sourcing.” DeepInfer is precisely the sort of technology that clause reaches — it is optimisation and scheduling know-how, the kind of thing the draft’s coverage of core private IP would otherwise capture. Publishing it under Apache 2.0 is not adjacent to the export-control question; it is a resolution of it. The licence is permissive in the strong sense: it grants a patent licence, it imposes no field-of-use restriction, it permits commercial serving, and it cannot be narrowed for anyone who already has a copy. A firm that wants its inference stack to operate outside a licensing desk has exactly one instrument that reliably achieves this, and it is a permissive licence.

Now consider the hardware asymmetry, which is where the release stops being purely a gift and starts being a position. Listing Ascend 950 alongside SM100 on day one, while omitting Hopper, is a statement about which silicon the software’s future is written against. A community that adopts DeepInfer to serve DeepSeek models has, without any decision on the merits, acquired a dependency on an inference stack whose second supported platform is a Chinese NPU. That is not a lock-in in the licence — Apache 2.0 forbids nothing — but it is a lock-in in the optimisation: kernels are tuned to architectures, and the architecture list is a design choice renewed with each release. The Reddit commenter who joked that a rival model “will delete the Ascend support” was making the point accidentally: support for a chip is a decision made by whoever maintains the code, and the licence’s permission to fork is not the same thing as the engineering capacity to maintain a fork.

Read against the 2025 document, the same release reads more soberly than as pure altruism. DeepSeek said in 2025 that the engine was too coupled to its internal infrastructure to release, and that maintaining another major project was beyond its team. In 2026 it released the part that is separable and kept the part that is not. That is an honest engineering judgement, and it is also the difference between contributing to a commons and publishing a component. The distinction carries institutional weight because a commons is defined not by the permissiveness of its licence but by whether the parties who depend on it can maintain it without the originator. A KV cache manager is not a glamorous omission. It is the part that determines whether a third party can run a service, and its absence means the recipient of DeepInfer is a user of DeepSeek’s design decisions rather than a co-maintainer of a shared engine.

There is a genuine and underrated achievement here that deserves to be named without qualification, because it is rare. DeepSeek published documentation of a decision reversal. The 2025 open-infra-index document stating why the engine would not be released remains in the repository, alongside the 2026 release that contradicts it in part. Firms usually do not leave their own reasoning where it can be checked against their later conduct; the combination makes the 2025 concerns falsifiable and shows exactly which of them were answered (maintainability, by narrowing scope) and which were not (internal coupling, which is precisely what the withheld harness is). That is a small piece of institutional hygiene and it is more than most open-source-adjacent corporate releases supply.

The checkable question. The test of whether DeepInfer is infrastructure or a component is not the licence field and not the download count. It is whether anyone maintains it besides DeepSeek. Concretely, over the next eighteen months: does a third party — an independent vendor, a cloud provider, or a hardware maker with a commercial interest in the alternative architecture — land substantive patches to DeepInfer’s scheduling, kernel or backend layers, and is Ascend performance parity achieved by someone other than Huawei and DeepSeek working together? A component receives contributions to its documentation and bug reports to its issues. An engine receives changes to its kernel selection and its distribution strategy from parties who do not work for the originator. Apache 2.0 permits both outcomes equally, which is why the licence field cannot answer the question — only the patch record can.

Sources:

🏗️ Institutional Change — A Hundred Billion Yuan and Two Hundred Shareholders: What the DeepSeek Round and the Moonshot Dispute Say About Corporate Form

1. The round. On 2026-10-22, Bloomberg reported that DeepSeek had raised about USD 15 billion, roughly RMB 100 billion — among the largest funding rounds ever completed in China. TechCrunch’s account notes the timing: the round closed two days after the 2026-10-20 US order directing federal civilian agencies to remove DeepSeek technology from their systems. Previous rounds: June 2026, more than USD 7 billion at a USD 50 billion valuation, closed with what Reuters described as an unusual deal structure; and in July, a reported effort to raise a further USD 1.5 billion at around USD 71 billion, with IPO plans for as early as the following year.

The Chinese-language reporting adds the structure, and the structure is the institutional content. Per 《财经》 (via ifeng, 2026-10-13), the round grew from an originally planned RMB 50 billion to about RMB 100 billion in under a fortnight, with new and existing shareholders taking roughly half each; the doubling did not raise the valuation, which remained at approximately RMB 500 billion pre-money. Lead new investors were predominantly industrial investors — Shixi Capital, SMIC Capital (中芯聚源), CPE, Legend Capital, Geely — while first-round holders CATL, Tencent, IDG Capital, Monolith and Shixiang all participated again, with CATL increasing its commitment to roughly RMB 10 billion from about RMB 5 billion in round one. Round one itself (opened April 2026, closed end-June) totalled about RMB 51 billion post-money near RMB 400 billion, with founder Liang Wenfeng personally contributing about RMB 20 billion, Tencent about RMB 10 billion, CATL entities about RMB 5 billion, and NetEase, JD, Monolith and IDG about RMB 3 billion each.

2. The pre-IPO engineering, which is the part that belongs in this series. Since late August 2026 DeepSeek has been engaging investment banks for a STAR Market (科创板) listing in Shanghai, has retained CITIC Securities, and plans to file as early as end-2026 with listing in 2027. The Chinese report states a detail that is easy to skim past and impossible to overstate institutionally: the sponsor is assisting in reviewing the backgrounds of investors, and has required some special-purpose funds to be restructured as blind pools (盲池备案), in order to avoid breaching the 200-shareholder ceiling after look-through (股东人数穿透后超过200人红线). Separately, Yan Wentao, formerly a partner at Hillhouse Venture Capital, joined as CFO in late September 2026.

3. The comparison case, in the same month. At Moonshot AI (月之暗面), the obstacle to a Hong Kong listing is not a shareholder count but a shareholder dispute. Reuters reported on 2026-10-11 that Moonshot executives were in talks to settle a conflict with a shareholder ahead of the planned IPO; Forbes’ account (2026-10-12) frames it around a USD 50 billion Hong Kong listing and an equity lawsuit, in which Yang and Zhang Yutao are accused of concealing their Moonshot equity stakes and of keeping Recurrent’s early venture backers in the dark about their new ownership roles while winding down their involvement in Recurrent. Moonshot’s valuation had reached USD 50 billion by early October, with an IPO targeted for early 2027. The counter-example for what a completed listing can look like: MiniMax, listed in Hong Kong on 2026-01-09, whose shares — per Caixin (2026-10-23) — peaked at HKD 1,330 on 2026-03-18 and have since fallen about 88% to HKD 160.7, below the IPO price, notwithstanding H1 revenue up 283.1% to USD 117 million and a gross margin of 17.9% against Zhipu AI’s 26.4%.

Institutional Analysis & Commentary:

Three firms, one month, three different constraints — and the differences are not about technology. They are about corporate form, which is the layer that determines what an organisation may do with its own success.

Start with the DeepSeek round’s most striking property, which the Chinese reporting states as a datum and which deserves to be read as a finding: the round doubled in size without moving the valuation. Fifty billion to a hundred billion renminbi, pre-money unchanged at about five hundred billion. A doubling of supply at a constant price is what happens when demand for a fixed quantity of a scarce thing exceeds it — but for a private company raising equity the more precise reading is that the constraint being relaxed was not the company’s need for capital but the investors’ need for allocation. DeepSeek’s controlling shareholder can choose how much of the company to sell; the doubling says the buyers wanted more than was on offer, and the price says the seller was not selling price. Which means the round is better understood as the construction of a shareholder register than as a financing.

That reading is confirmed by the detail about blind pools. The 200-shareholder ceiling after look-through is not a formality; it is the hard boundary of the Chinese corporate form that precedes a public listing, and it is enforced by counting through fund structures to the natural persons behind them. A company that has raised, across two rounds, from founder, strategics, industrial funds, consumer-internet platforms and battery makers must eventually present a countable register. Requiring special-purpose vehicles to convert to blind-pool filings is not paperwork; it is the deliberate erasure of information about who is ultimately behind a holding, performed for the purpose of fitting under a statutory limit. The alternative reading — that blind pools conceal something — is not required and not supported; the disclosed purpose is arithmetic. But note what the arithmetic produces: a corporate form in which the identity of the marginal shareholder is, at the moment of listing, deliberately unobservable, and in which the sponsor’s due diligence on investor backgrounds is conducted so that the count will pass. That is a governance fact about the pre-IPO stage in this market, and it is the kind of fact that a mature disclosure regime would treat as a tension rather than a procedure.

Now stack the three forms. DeepSeek is solving a counting problem: too many beneficial owners for the domestic listing route, so the register must be compressed. Moonshot is solving a title problem: a shareholder’s claim on the equity is contested, and a contested title is disqualifying for a listing in a way that an unfavourable price is not — which is why the settlement talks are a listing prerequisite and the USD 50 billion valuation is not. MiniMax has already solved both and is now discovering a third problem, which is repricing: an 88% drawdown from a March peak while revenue grew 283% is the market saying that the price paid at IPO was a statement about a growth trajectory, not about a business, and that the trajectory has been re-underwritten as competition compressed margins to 17.9%.

The institutional lesson in the three taken together is not “Chinese AI is overvalued” — that is a claim about prices, and this series does not make price claims. It is that the Chinese AI sector is in the middle of a transition from an ownership regime organised around private agreement to one organised around public disclosure, and the friction is concentrated in exactly the two places where a private agreement is not designed to be inspected: the shareholder count and the shareholder title. Neither of those is a technology problem. Both are reasons why the form of a company — how many people hold it, and whether their holdings are clean — becomes a binding constraint at precisely the moment the technology becomes valuable enough to list.

There is a counterweight that a fair reading requires. A shareholder cap is a consumer-protection instrument, not an obstacle course. The 200-person limit exists to prevent a company from raising from the public without being subject to public-company obligations; compressing the register to fit under it is compliance, not evasion, and blind-pool structuring is a recognised vehicle used in every market with look-through rules. Similarly, a shareholder dispute ahead of listing is normal corporate practice worldwide, and the existence of settlement talks is evidence the parties prefer the listing to the litigation. The accurate observation is therefore narrower than a critique and more useful: this sector’s constraint has shifted from whether it can build the model to whether it can fit its cap table into a form that a public market can accept. That shift is a sign of maturation, and it is also the moment at which the incentives governing a lab’s decisions change — a company preparing to list optimises differently from a company preparing to fight.

The checkable question. What will settle whether the round is an endorsement or a bridge is not the price but what the filing discloses. When DeepSeek’s STAR Market prospectus appears, three things in it are testable against what has already been reported: does the disclosed investor list survive look-through without the blind-pool compression, or does the compression persist into the filing; does the disclosed use of proceeds name compute, chips and inference capacity as the reported infrastructure expansion implies; and does the filing attribute revenue at a scale consistent with the reported USD 1 billion annualised figure, with a gross margin that says whether open-weight releases and API pricing are compatible? A round that doubled at a constant valuation because investors wanted allocation is a story about scarcity. A prospectus that shows where the money went is a story about strategy. The gap between the two is where the real information about this sector’s institutional direction will be found.

Sources:

📡 WeChat Monitor — A Foundation Goes to Campus, and the Question a Campus Visit Does Not Answer

1. The event, from the foundation’s own account. On 2026-10-21, the WeChat account of the Tiangong Kaiwu Open Source Foundation (天工开物开源基金会) published a report on “Open Source Campus Tour · Shandong University Station” (开源高校行·山东大学站), held at Shandong University’s Software Park campus and jointly hosted by Shandong University and the foundation, with Modelers Community (魔乐社区) and Ant Open Source as ecosystem partners. The theme was “Open Source Sets Sail · Intelligence Creates the Future — Open Source Technology and Youth Practice in the AI Era” (开源启航·智创未来——AI时代的开源技术与青年实践).

2. The scale and the format, precisely as reported. More than 400 students attended, split across three classrooms. Four speakers presented, and the list is a clean cross-section of the sponsor ecology: Zhuang Biaowei (庄表伟), deputy secretary-general of the Tiangong Kaiwu Open Source Foundation, on “the value and system of the open-source ecosystem,” analysing how open source creates technical, economic and social value across three layers — product, community, ecosystem; Yan Shu (闫舒), open-source and R&D efficiency manager at Inspur General Software, on UBML, exploring intelligent low-code development via a metamodel architecture, from code generation, explanation and test assistance to domain-model-based application construction; Dong Chao (董超), open-source governance expert at Ant Group, on “from code to community — open source in the AI era,” walking through the path of a single contribution — discovering a problem, discussing a solution, submitting a change, review and verification, and continuous maintenance — and introducing Ant Group’s open-source light training camp (开源轻训营); and Zhang Shufeng (张树峰), product lead at Modelers Community, on building a domestic AI ecosystem through open collaboration, covering the community’s resources and practice in model, dataset and tool hosting and application collaboration.

Institutional Analysis & Commentary:

The report is a campus event and it should be read as one — four talks, 400 students, three classrooms. What makes it legible to an institutional series is the composition of the speaker list, because a roster is a document about who bears the cost of recruitment.

Four speakers, four different kinds of institutional actor, and four different things each of them needs from the room. The foundation supplies the frame — the product/community/ecosystem三层 analysis is the pedagogical scaffolding, and a foundation’s interest in a campus is durable and generic: it needs a pipeline of future contributors to the projects it hosts, and it has no product to sell. The vendor (Inspur) supplies a technology talk, and its interest is specific: UBML is a project, students who learn it are future users and possibly contributors, and a metamodel low-code framework needs exactly the kind of practitioner awareness a university audience can supply. The platform company (Ant Group) supplies the process talk — how a contribution actually travels from problem to maintained code — plus an entry point, the light training camp, which converts interest into a structured first contribution. The hosting community (Modelers) supplies the infrastructure talk, because a community whose value is model, dataset and tool hosting needs users who know it exists.

This is a division of labour, and it is worth naming what each party contributes and what each receives, because the arrangement is more coherent than the usual critique of campus outreach assumes. The foundation contributes legitimacy and curriculum; the vendor contributes a real codebase; the platform company contributes a governed contribution path; the community contributes hosting. In exchange, each receives access to the same scarce resource — attention, before it has been allocated to something else. That is not a distortion of open source. It is close to how contributor pipelines are built everywhere: a first contribution is a threshold, and thresholds are crossed with help.

The institutional question the report does not answer is the one that determines whether 400 attendees become 4 contributors. Open source’s contribution mechanism is unusual among professional activities in that it requires no credential, no interview and no permission — the licence is the only entry condition, and a stranger with a patch is a peer. But a campus event is a presentation format, and presentation formats are structurally weak at producing the one thing that matters: the first accepted pull request. Dong Chao’s talk, on the evidence of the description, is the only one aimed at that threshold — problem, discussion, change, review, maintenance — and it is paired with a training camp, which is the mechanism that converts a talk into an attempt. The remaining three are orientation: what the ecosystem is, what a project is, what a platform is. Orientation is necessary and it is not sufficient, and the ratio here — one threshold-crossing session to three orientation sessions — is the detail a reader should weigh.

There is a second observation, more specific to this foundation and worth stating carefully because it is a description rather than a judgement. Tiangong Kaiwu is a foundation, and foundations in this ecosystem occupy a particular position: they hold assets and projects, and they must demonstrate public benefit to sustain the mandate that lets them hold them. Campus tours are an efficient way to demonstrate public benefit — they are visible, they are educational, they cost little, and they produce photographs of full rooms. Whether they produce contributors is a different question and a slower one. The measure that would distinguish the two is not attendance but retention: of the 400 students in those three classrooms, how many have an accepted contribution in a Tiangong Kaiwu-hosted project within twelve months, and does the foundation track it? A foundation that tracks its campus programme by attendance is running an outreach function. A foundation that tracks it by merged patches is running a contributor pipeline. The report provides the first number and not the second, which is normal, and which is also the reason the first number is the one that gets published.

The checkable question. Straightforward and public: does the Shandong University cohort appear in a project’s contribution record? Concretely — are there commits, issues or pull requests from Shandong University-affiliated accounts in Tiangong Kaiwu-hosted projects or in the Modelers/Ant repositories presented at the event in the twelve months following 2026-10-21, and does Ant Group’s 开源轻训营 publish cohort-to-contribution conversion figures? Where open source is concerned, the only honest measure of a recruitment event is what the recruit did afterwards, and — unlike attendance — it is a record that anyone can read.

Sources:

🔍 Commentary — Three Ways to Make Publishing Profitable, and the One That Was Legislated

Three documents from eleven days in October 2026 describe a single institutional shift, and none of them mentions the others.

On 2026-10-16 two Chinese ministries published a draft export-control catalogue containing twenty-three AI technologies, switched on a registration system the same day, and wrote into the text that technology already in the public domain by means of open sourcing is not within the scope of management. On 2026-10-20 DeepSeek published its production inference engine under Apache 2.0 and listed Huawei’s Ascend 950 next to Nvidia’s SM100 as a first-class platform. On 2026-10-22 it was reported to have raised about RMB 100 billion — a round that had doubled in a fortnight without moving the valuation, in the middle of compressing its shareholder register to fit under a two-hundred-person ceiling ahead of a Shanghai listing.

The mechanism connecting them is a price, and it is a price that the state has just altered. Consider what it costs a Chinese AI firm to keep a frontier model closed. It must licence it abroad — which, after Announcement 48, means a registration entry, a licensing desk and an undefined “automatic” process whose own Q&A hedges with the word “basically.” It must accept that its decision is now on a record, filed before the fact, in a system that also receives its technical specifications. And it must accept that the rule governing all this is a draft whose final form will be set by ministries that have already been advised, in a substantial academic literature, to build a negative list barring the release of certain weights altogether. Keeping a model closed is now an act of filing.

Compare the other option. Publish the weights, and the technology leaves the regime entirely — not expedited, not exempted, but outside scope, permanently, because a published weight cannot be recalled. Publish the inference engine under Apache 2.0, and the same applies to the operational know-how. The cost of publishing is the loss of pricing power over something the world now has. The benefit is that the thing becomes unregulable by anyone, including the state that produced the firm. Between a filing desk and a permanent exit, a firm optimising over a five-year horizon does not deliberate long, and the doubled-without-repricing capital round is what a balance sheet looks like when that calculation has been made.

Now, the important qualification, because a reading that stops here would be a reading of the control regime as an open-source policy and it is not one. The exemption exists because it serves the state’s interest, not because it serves open source’s. Read the announcement for what it is trying to protect and the exemption becomes obvious. Section one of the AI coverage exempts not only published technology but also the provision of API calls, app and web interaction, SaaS and on-device deployment to foreign users. What remains controlled is the transfer of capability and know-how — weights, source, chip design, cluster technology. That is a regime designed to permit the export of services while retaining leverage over the export of means. A published model, on this logic, is not a loss of a national asset; it is the cheapest possible mechanism of standards diffusion, because a party who builds on your architecture inherits your toolchain, and a party who inherits your toolchain argues in your ecosystem’s terms. The Chinese academic formulation that Jamestown quotes says it without embarrassment: “using technology to carry rules” (以技术带规则), described as “more sustainable and more persuasive than exporting rules on their own.”

That is the point at which this must be named plainly, because it is the exact inversion of what open source is for. A licence is a grant of rights to strangers; it is not a delivery mechanism for anyone’s standards, and the moment publishing is understood as a channel through which norms travel, the commons has been conscripted. The distinction is not academic. A commons survives on reciprocity — on contributions flowing back, on maintainers who are accountable to the people who depend on them, on the boundary between provision and appropriation being visible. A diffusion strategy survives on asymmetry — the publisher keeps the harness, the recipient gets the engine, and the fact that the engine is free is what makes the asymmetry acceptable. DeepInfer makes the distinction concrete in a single repository: Apache 2.0 on the scheduling and kernels, nothing on the KV cache manager, cross-machine stage separation or replica load balancing. The licence is the widest possible grant. The stack is deliberately incomplete. Both facts are true, and only the pair of them describes the strategy.

There is a counterweight in the same fortnight that a fair reading has to include, and it is not small. Fifteen billion dollars and a production inference stack published under a permissive licence, from the same firm, in the same week, is a real and unusual contribution to a global technical commons — one that a great many developers in a great many jurisdictions will use to run models more cheaply than they otherwise could, with no permission step, no registration and no relationship to the supplier. Whatever the strategic reading, the artefact is an artefact, and it can be forked by anyone who disputes the strategy. That is how a commons absorbs the ambitions of its contributors: not by preventing the ambition, but by making its output divisible from it.

The final observation concerns the firm rather than the state, and it is the one this series will keep returning to. DeepSeek in October 2026 is doing three things at once: raising the largest round in its sector, publishing its operating stack, and preparing to list — which means converting an ownership structure built from private agreements into one that a public market will accept. Those three run on different clocks. Publishing moves at the speed of a git push and is irreversible. Raising moves at the speed of a shareholders’ agreement and is reversible until it closes. Listing moves at the speed of a regulator’s look-through and is reversible until it prices. A firm that publishes irreversibly while preparing to disclose fully is a firm whose strategy has already been fixed by the thing that cannot be undone. The prospectus, when it comes, will describe a company whose most valuable assets have been given away under a licence it cannot revoke. Reading that document will be the moment the relationship between Chinese open source and Chinese AI capital becomes legible — not because the filing will explain it, but because a filing is required to state what a company owns, and the interesting question is how it accounts for what it has published.

The checkable question. One instrument decides how all of this is remembered: the final text of the revised Catalogue, and the implementing guidance that defines “already entered the public domain.” If the definition is generous — permissive or restricted-use weight releases, plus published training code — then open sourcing remains the dominant strategy for any Chinese lab whose model crosses the thresholds, the exemption becomes the most consequential piece of open-source policy any state has issued, and the sector’s competitive dynamic shifts decisively toward publishing. If the definition is narrow, the exemption becomes a transitional artefact of a drafting process and the negative-list literature wins by implementation rather than by argument. The draft’s comment period closes 2026-11-15. Everything in this briefing is a reading of a document that has a month to change.

Sources: