⚠️ 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 — 2026-10-30

🏛️ Policy & Regulation — “Openness Is the Precondition”: Reading the Inclusive Plan’s Two-Year Progress Report as an Institutional Document

1. What was published, and by whom. On 2026-10-24, at the 2026 Global Governance Forum (Xiongan) (2026全球治理论坛·雄安), Foreign Minister Wang Yi released the Progress Report on the AI Capacity Building Inclusive Plan (《〈人工智能能力建设普惠计划〉进展报告》) — the seventh of eight concrete actions he itemised in his keynote, alongside APEC, UN reform, the Global Governance Friends Group, the BBNJ agreement, human-rights governance and outer-space governance. The report is issued by the Global AI Innovation Governance Centre (全球人工智能创新治理中心), dated October 2026, seventeen pages, and its PDF sits on the Ministry of Foreign Affairs domain. The 2024-09 origin is restated verbatim: five visions — infrastructure connectivity, “AI+” industry empowerment, literacy and talent, data security and diversity, safety and controllability — organised into ten actions.

2. The structural fact, before any argument about content. Read the table of contents and count. Section 03, the ten actions, runs: (1) North-South, South-South and trilateral cooperation; (2) model R&D and industry empowerment; (3) green development, climate and biodiversity; (4) building international cooperation platforms and an open-source ecosystem; (5) short- and medium-term education and training; (6) human-resource assistance and seminars; (7) public literacy and digital rights for women and children; (8) corpus construction and linguistic diversity; (9) data governance; (10) strategic policy exchange on safety governance. Each action is documented with one numbered case. The cases, in order: FlagOS; Huawei and the South African Passenger Rail Agency (PRASA); the MAZU early-warning scheme of the China Meteorological Administration; the World AI Cooperation Organisation (WAICO); the Panshi scientific foundation model; the AI capacity-building seminar series; the APEC women-friendly AI toolkit; the Lao-language model built by the Guangxi Data Bureau with China-ASEAN Information Harbour; the Sangfor security operations centre at Srinakharinwirot University, Thailand; and the BRICS leaders’ statement on AI global governance. Exactly one of the ten cases is an open-source artefact, and it is not a model — it is a system software stack.

3. The four lessons, and the fourth is the document’s thesis. Section 04 states: (1) technical access is the starting point, autonomous development is the goal — with the limit conceded immediately, that open models, cloud services and platforms “cannot replace local institutions’ capacity for decision-making, R&D and governance”; (2) demand pull and joint consultation determine whether a project lands, including the explicit warning against projects that are built and never used (建而不用、用而难续); (3) data, models, talent and governance must advance together; and (4) “开源开放是实现普惠发展的前提” — open source and openness are the precondition for inclusive development.

That fourth sentence deserves quoting in full, because of what follows it. The justification is not technical but political-economy: if AI tools “are closed and monopolised by a small number of countries or enterprises, it will only further widen the North-South gap”; only persistent openness “lowers the threshold of innovation” and lets open results “benefit the Global South, so that the right to development is effectively guaranteed in the intelligentisation process.” Then the qualifier, which is the load-bearing half:

“因此，开放合作需要配套的知识框架，这当中包括亟待建设的工具链、测试环境和相关支持。” — Therefore, open cooperation requires a supporting knowledge framework, including toolchains, test environments and related support that remain to be built.

4. What the report counts, and how it counts. The gap section rests on World Bank figures — high-income countries hold 77% of hosted data-centre capacity, 87% of AI models and 91% of venture investment; Africa holds under 1% of data-centre capacity and fewer than 1,000 chips usable for training AI models — and on UNCTAD’s 2025 Technology and Innovation Report: fewer than one third of developing countries have an AI strategy, and 118 countries did not effectively participate in major AI-governance discussions. The multilateral record is given as: UNGA resolution A/RES/78/311, adopted by consensus in July 2024 with 143 co-sponsors; the Friends Group’s first meeting on 2024-12-03 with 80-plus countries and UN bodies, and a 2026-05-05 thematic session with 50-plus countries and 120-plus delegates; WAICO signed at Shanghai on 2026-07-16 by 29 states, now holding 38 founding members after Brunei, Vietnam, Georgia and Tanzania acceded, headquartered in Shanghai, and described in the report as the first intergovernmental organisation in the AI field; and a pledge of 5,000 AI training slots for developing countries over five years, announced by President Xi at the July 2026 World AI Conference.

5. The one case that is substantively about open source. The FlagOS entry, quoted from the report rather than paraphrased: led by BAAI with Peking University, Tsinghua, the Institute of Computing Technology at CAS and more than ten chip vendors; goal “develop once, run on many chips”; supports 21 chip vendors; FlagOS 2.0 adds six domain operator libraries for a total of 1,454 operators, described as the world’s largest multi-chip operator library; supports multi-chip deployment for over 90% of mainstream open models; 200,000 users worldwide; and received an ITU AI for Good outstanding-case award at the Geneva summit in July 2026. Everything in that entry is checkable in the technical sense — operator counts, vendor counts, model coverage. None of it is a governance claim.

6. A figure that has to be recorded, because this series has already printed three versions of it. The report’s companion coverage repeats Qwen’s “30 billion-plus cumulative global downloads.” The chain behind that number, verified for this briefing: on 2026-07-20 at a State Council Information Office press conference, MIIT chief engineer Wang Weiming said the cumulative global downloads of China’s open-source large models had surpassed 10 billion (突破100亿次); the MOFCOM think tank article of 2026-10-13 gives Qwen at 3 billion (30亿), against which China Daily’s English edition reports Hugging Face’s 2026 count for Qwen as 2.045 billion and for Google 418 million and Meta 227 million; and The Next Web notes that the 3-billion figure originated in an Alibaba statement emailed to reporters, while Hugging Face’s State of Open Models (published the day earlier) counted 2.045 billion on its own Hub, or 2,061 million including every repository, and 151,448 Qwen derivatives against Alibaba’s claimed 300,000-plus. This series’ 2026-10-06 briefing read the MIIT number as “100 billion.” That reading was wrong by a factor of ten. The paragraph below is about how such figures are constructed, not about correcting a number.

Institutional Analysis & Commentary:

Begin where the arithmetic is cleanest, because the arithmetic is the day’s most usable lesson in institutional method.

Alibaba says 3 billion downloads. Hugging Face counted 2.045 billion on its Hub in 2026, 2.061 billion including repositories without declared parameter counts. Both can be true: Hugging Face measures its own Hub, and Qwen also ships through ModelScope. The gap that is harder to reconcile is the derivatives — 151,448 on the Hub against a claimed 300,000-plus. The Next Web calls that roughly a 148,000-model difference and says Alibaba’s number came from an emailed statement rather than a published table. Neither party is lying, and neither number is a finding. One is a sensor reading with a stated aperture; the other is a counter that has not published its aperture. The institutional point is exactly this: a metric without a stated denominator and a stated window is not evidence, it is a claim about a market. This series has spent a year criticising Chinese grading instruments for exactly the same move — a tier, a score, a “global first” with no published methodology — and the discipline has to run in both directions or it is not a discipline. The Hugging Face report, whatever its own editorial choices, publishes its aperture: repositories with declared parameter counts, January to August 2026, Hub only. That is the difference between a measurement and a statement, and it costs nothing to make.

Now the report itself, and the honest reading is not the cynical one.

The Inclusive Plan’s centrality claim is “openness is the precondition for inclusive development.” That is a logical operator, not a component: it says without openness, inclusion does not occur. Set it beside this series’ 2026-10-28 file and the mechanism becomes concrete — a German ministry’s eligibility criterion could not be satisfied by improving an artefact, while the Apache 2.0 licence could be read by anyone with a browser; the parties with the least capacity to relocate a legal person were also the parties least able to enter the argument at all. Openness as precondition is the same finding, stated from the receiving end. A country that cannot obtain the weights cannot fine-tune, cannot evaluate, cannot audit, cannot localise, cannot refuse. That is not a slogan; it is a supply-chain fact, and the report is right about it.

But a claim in a foreign-ministry progress report is a different institutional object from the same claim in a licence, and the report’s own ten-action structure makes the point for it. Of ten actions, one is nominally about open source, and its documented case is a software stack for running models across many chips — not a licence, not a repository, not a contribution process, not a dispute clause. The other nine are platforms, deployments, seminars, toolkits and statements. This is not a failure of sincerity; it is a category boundary. A progress report on international cooperation reports cooperation, and cooperation is counted in participants, countries and installations because those are the things a ministry can verify, attribute and report upward. A cooperation output and a contribution artefact are different institutional forms, and the difference is not scale: one is created by signature, the other by strangers deciding that their effort will not be wasted.

Which is why the most important sentence in seventeen pages is the qualifier, not the thesis. Toolchains, test environments and related support that remain to be built. That sentence is the author knowing precisely where the constraint sits — and it is the sentence with no case study attached, in a document where every other claim has one. A test environment is not a deliverable a ministry announces; it is accumulated calibration that a strange developer trusts enough to depend on. The report lists it as pending, in the same section where it calls openness the precondition. The gap between those two observations is the whole institutional distance, and the report states it more plainly than most external critics would.

And then the document from the same month that proves the point from the opposite direction, at the mundane level where these things actually bite. On 2026-10-28, the Central Government Procurement Center published the award notice for its 2026 Windows operating-system framework solicitation — and the machine-readable metadata carries packageName, purchaseProjectName and supplierName empty, with the notice text recording that there was no effective supplier (无有效供应商), on the ground that Windows and Linux desktops were not substitutable for the covered purchasers. Read the two documents as a pair. The Plan argues that closed artefacts widen the gap; the notice records what a closed artefact’s procurement looks like when the vendor does not bid — a properly formatted document with three blank fields. Nobody behaved badly in either case. The ministry argued a correct thesis; the procurement centre dutifully published a null result. Together they say one thing: access is not a property of a licence, and neither is produced by a cooperation programme.

Now the part that has to cut inward, or this is only commentary. This series’ files of 2026-10-28 and 2026-10-29 documented two state instruments in the same fortnight: a German eligibility rule about where a vendor is incorporated — an entity test, satisfiable only by relocating a legal person — and a Chinese procurement standard whose starred item 1 is a gate on same-source compatibility across five CPU architecture families, with item 20 exempting non-Linux kernels from a Linux-specific requirement — an artefact test, satisfiable by building, written in the incumbent’s vocabulary. Today’s report supplies the third position: a state arguing in a foreign-ministry document that openness is the precondition, and reporting on platforms rather than artefacts. Three instruments, three relationships to the artefact, one question underneath all of them: who may obtain the thing, and on what terms. Only one of the three can be answered by engineering.

The honest counterweights, because the easy version of this critique is wrong twice. First: the report does not overclaim, and its first lesson concedes its own limit — that open models cannot substitute for local institutional capacity for decision-making and governance. That is exactly the distinction this series keeps drawing between project availability and institutional capacity, written into an official document as a limitation. Second: FlagOS is a real artefact with real numbers, and its placement in the case list is not decorative — multi-chip portability is the binding constraint for any country that cannot buy one vendor’s accelerator stack. Twenty-one vendors and 1,454 operators is a working answer to a hard problem, and it earned an award from a UN agency it does not control. Third, and least comfortable: the ten-action structure is not a Chinese peculiarity. Every development-cooperation programme in every capital reports in events and participants, because a programme is what a foreign ministry administers. Even the report’s weakest-sounding item — the toolkit, the seminar cohort, the statement — is a genuine transfer of policy vocabulary, and policy vocabulary is the input to domestic rule-making.

The checkable question — three observables, each answerable from a public record. First: does the 2027 progress report contain, alongside its platforms and seminars, one entry that is a repository — a named licence, a contribution guide, a review history, and a maintainer who is not the original vendor? One such entry means the fourth lesson was operationalised into something a stranger can join. Ten more platforms means the thesis is being applied to other people’s artefacts. Second: does any of the ten cases already reported acquire an open artefact within a year? A warning system deployed in eight countries and a toolkit published in five languages are outputs. A repository with outside committers and a published review policy is an institution, and the two differ by whether a person who has never met the maintainers can rely on the outcome. Third, and the one that generalises: when this series cites a download or derivative figure from any source — Chinese ministry, Alibaba, Hugging Face — does the source publish its aperture, its window and its denominator? The 2.045-billion-versus-3-billion gap is not a scandal; it is the ordinary condition of an ecosystem measured by platforms that do not see each other’s traffic. But a briefing that reports one number without the other is doing the same thing it criticises. The measure of an open ecosystem is not the number it publishes. It is whether a stranger can reproduce it.

Sources:

🔍 Commentary — A Precondition Is Not a Portfolio

The sentence is one character away from an ordinary diplomatic formula, and it is not one: openness is the precondition for inclusive development. Precondition — not contribution, not component, not one of several viable paths. Whoever wrote section 04(4) of the Inclusive Plan’s progress report chose a logical operator, and a logical operator is the kind of claim that can be tested.

So test it against the document’s own inventory, which is the fair test and the only one available.

Ten actions, ten cases. FlagOS: a genuine open-source system software stack — 21 chip vendors, 1,454 operators, 200,000 users, an ITU award. WAICO: an intergovernmental organisation, 38 founding states, headquartered in Shanghai. Huawei with the South African rail operator: a commercial deployment with a measured efficiency gain. MAZU: an early-warning scheme in eight countries. Panshi: a scientific foundation model opened to 96 Belt-and-Road science alliance members. The seminar series: four cohorts, 60-plus countries, 100-plus officials. The APEC toolkit: a document. The Lao-language model: a deployment, with its technical framework extended to three more ASEAN states. Sangfor: a commercial security-operations centre. The BRICS statement: a statement.

One of the ten is an artefact whose licence you can read. That is not a scandal — it is a category boundary, and naming it precisely is the job here. A progress report on international cooperation reports cooperation, and cooperation is counted in participants and installations because those are what a ministry can verify. The artefact — the thing with a licence file, a commit history, a review protocol and an exit right — is governed by an entirely different apparatus, and that apparatus does not appear in this document at all.

Here is the asymmetry that makes it matter. A cooperation platform can be created by signature: WAICO went from signing ceremony to 38 founding members inside four months. A contribution process cannot be created by signature, because it is not a body — it is a set of expectations a stranger can rely on before investing months of work. Licences, review norms, disclosure obligations, dispute resolution, the right to fork: these accumulate slowly, they are written mostly by people who have been burned, and they are the reason a developer in Jakarta or Lagos can commit code to a project whose maintainers they have never met and know what happens next.

Which is why the most important sentence in seventeen pages is the qualifier, not the thesis. After asserting that openness is the precondition, the report writes: therefore open cooperation requires a supporting knowledge framework, including toolchains, test environments and related support that remain to be built. That is the author knowing exactly where the constraint sits — and it is the only claim in the document with no case study attached, in a document where every other claim has one. A test environment is not a deliverable a ministry announces. It is accumulated calibration that a stranger trusts enough to depend on. The report names it as pending, four pages after calling openness the precondition, and that distance — not any external criticism — is the most accurate available statement of where the work is.

And a document from four days later makes the same point where these things actually bite. On 2026-10-28 the central procurement centre published the award notice for its Windows operating-system framework solicitation. The machine-readable metadata has packageName, purchaseProjectName and supplierName empty, and the notice text records the reason: no effective supplier, because Windows and Linux desktops are not substitutable for the covered purchasers. Read the pair together and they are one document. The Plan argues that closed artefacts widen the gap; the notice records what a closed artefact’s procurement looks like when its vendor does not bid — a properly formatted document with three blank fields. Nobody behaved badly in either. The ministry argued a correct thesis; the procurement centre dutifully published a null result. Together they say: access is not a property of a licence, and neither one is produced by a cooperation programme.

Now the part that has to cut inward, or this is only commentary. This series’ last two files documented two state instruments in the same fortnight: a German eligibility rule about where a vendor is incorporated — an entity test, satisfiable only by relocating a legal person — and a Chinese procurement standard whose starred item 1 gates on same-source compatibility across five CPU architecture families, with item 20 exempting non-Linux kernels from a Linux-specific requirement — an artefact test, satisfiable by building, written in the incumbent’s vocabulary. Today’s report supplies the third position: a state arguing in a foreign-ministry document that openness is the precondition, and reporting on platforms rather than artefacts. Three instruments, three relationships to the artefact, one question underneath all three: who may obtain the thing, and on what terms. Only one of them can be answered by engineering. Naming that is not scoring points for anyone; it is the only way a party outside all three processes knows whether to hire lawyers or engineers.

And a number in this file that belongs to me, not to the sources. This series’ 2026-10-06 briefing read the MIIT download figure as “100 billion.” The correct reading is 10 billion — 突破100亿次 is 10 billion, and Global Times’ English edition of the same 2026-07-20 briefing renders it exactly that way. The error was mine, it was a factor of ten, and it sat in an argued paragraph for three weeks. It matters institutionally for a reason that has nothing to do with embarrassment: this briefing’s whole method is that claims should state their aperture and be checkable, and a figure that travels without its denominator is the exact failure mode it criticises in other people’s instruments. 2.045 billion on the Hub, 3 billion claimed cross-platform including ModelScope, 151,448 derivatives counted against 300,000-plus claimed — none of those numbers is wrong, and none of them is evidence until someone says what it counts. A metric is an institution. It has rules, and the rules are the point.

The checkable question, and it is deliberately small. First: does the 2027 progress report contain, alongside its platforms and its seminars, one entry that is a repository — a named licence, a contribution guide, a review history, and a maintainer who is not the original vendor? One such entry would mean the fourth lesson was operationalised into something a stranger can join. Ten more platforms would mean the thesis about openness is being applied to other people’s artefacts while the state’s own instruments keep selecting for portfolios and purchasers. Second: does any of the ten reported cases acquire an open artefact within a year? A warning system deployed in eight countries and a toolkit published in five languages are outputs. A repository with outside committers is an institution, and the difference between them is precisely whether someone who has never met the maintainers can rely on the outcome.

A precondition is a claim about causation. A portfolio is a list of what you have. The distance between the two is not dishonesty — it is the whole distance between having a policy and having an institution, and it is measured in years, not in pages.