⚠️ 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 — 2026-11-16
🏛️ Policy & Regulation — Three Hundred Million Yuan for a Commons That Is Not One: BDA’s Open-Source Subsidy Schedule
1. The document, and who issued it. On 2026-11-02 the Beijing Economic-Technological Development Area (BDA, 北京经济技术开发区) — Yizhuang — issued 《北京经济技术开发区关于加快建设开源开放信创产业高地的若干措施》 (Jing Ji Guan Fa [2026] No. 16), published 2026-11-05, effective on publication and valid until 2029-12-31, replacing the 2025 version (No. 3). It is issued by the Administrative Committee of the zone, drafted by its Information Technology Industry Bureau (信息技术产业局), and applies to all entities operating lawfully within the 225 km² of Yizhuang New Town; the interpretation note lists exactly which clauses were revised — one, three, four, five, seven and eight. The stated purpose is to make the National Xìnchuàng Park (国家信创园) a national high ground for open-source, open information-technology application innovation (开源开放信息技术应用创新).
2. The ten actions, and the money attached to each. The document is structured as ten named actions. The ones with figures: the goose-formation plan (雁阵计划) — up to CNY 100 million per year for up to five years for a lead enterprise that pulls the chain, on top of R&D support capped at 50% of total project cost; technology foundations — matching support up to 1:1 for national or municipal major research tasks, capped at CNY 30 million per project per year; scenario leap — up to CNY 30 million for flagship applications; platform empowerment — up to CNY 30 million per year for up to three years for operating an industry lab, security-assessment body, hardware–software adaptation and test centre, industry-education base or shared-services platform.
3. The fifth action, which is the one this entry is about. Open-source aggregation (开源聚力行动), quoted in full because every clause is a subsidy line: support open-source communities, platforms and projects with independent core technology; build out open-source code hosting, code inspection and critical open-source component backup infrastructure; build a national artificial-intelligence open-source community (国家人工智能开源社区), support open-source large models and toolchains to move in (入驻), build an open-source ecosystem operations platform, and cultivate an open-source large-model community with global influence. Then the four payments: up to CNY 1 million one-off for a high-quality open-source project; up to CNY 3 million one-off for donating a high-quality open-source software project to a national-level open-source organisation and having it accepted (捐赠给国家级开源组织并被接收); up to CNY 1 million for a startup or enterprise building a commercial product on an open-source project, plus “move-in-with-your-bag” office space; and up to CNY 3 million, at 50% of actual expenditure, for hosting a national open-source ecosystem event. The standards action adds up to CNY 1 million to the lead or principal drafting unit of an implemented standard.
4. The targets, and the adjective. By 2027: 1,000 xìnchuàng enterprises, revenue above CNY 100 billion, 50 flagship solutions, and at least 50 high-quality open-source projects; by 2030: a national open-source and xìnchuàng high ground with global influence. And one phrase that does not appear in the version of the policy summarised by most outlets but is unambiguous in the 2025 text of the same action and its 2026 revision: the community to be built is a software- and hardware-sovereign (软硬件自主可控) national AI open-source community. In the 2026 text the sovereignty qualifier has moved into the enclosing clause — the park is to be 自主可控、安全可靠、竞争力强 — and the community is named national.
Institutional Analysis & Commentary:
Six weeks of this series has argued a single asymmetry: every instrument in the Chinese open-source ecosystem is fitted to a transaction with a buyer, and nothing is fitted to the transaction with a stranger. Today’s document is the first that appears to contradict the thesis, and the way it contradicts it is more informative than the thesis was.
Start with what is genuinely new, because credit should be given precisely. Every subsidy examined so far paid for use: the adaptation voucher paid an SME to make a device run the system; the certification paid a body to tell a candidate that they know it; the curriculum paid a university to teach it; the procurement contract paid a consortium to deliver it. This policy contains a line that pays for giving something away. “Donate a high-quality open-source software project to a national-level open-source organisation and have it accepted — up to CNY 3 million.” That is a price on the act of relinquishing control, and it is the first such price in this series. Read alongside the university clause from yesterday’s 2026-11-14 entry — eight hundred courses and no course in which a student must get a patch merged — the donation subsidy is the closest thing the ecosystem has yet produced to a budget line for the outbound act of a commons: the moment when an asset stops being someone’s product and becomes something others can build on. In the openEuler history (2026-11-14), that moment was Huawei’s November 2021 donation to OpenAtom, and it happened without any subsidy. A city now thinks it needs to pay for it, which is itself a finding: the donation step is not self-evidently attractive to the firms that own the projects.
Second, and now the harder reading, which is the entry’s real subject: look at what is not in the schedule. Search the ten actions for a payment that goes to a person. There is exactly one clause about humans, and it is the talent action — recruitment of senior researchers, the Yicheng talent designation, an industry-AI institute, joint course and textbook development with universities, student internships. Every figure in the document is denominated in enterprises, platforms, projects, standards or events. Not one yuan is attached to a maintainer, a reviewer, a committer, or an answer to an issue at three in the morning. The largest single number in the open-source action — CNY 3 million — is paid for donating a project to an organisation and for hosting a conference. The two things that cost money and produce a commons — merging a stranger’s patch, and keeping a repository’s security advisories current — are not line items, because they are not transactions with a legal person that an administrative committee can disburse against. This is the same structural constraint this series identified on 2026-11-11 and 2026-11-15, arriving now in a fiscal document rather than a standards committee: the fiscal instrument can only see legal persons, and a commons is built out of acts by natural ones.
Third, the qualification in the donation clause, which is the whole analysis in five characters. “Donated to a national-level open-source organisation (国家级开源组织) and accepted.” Two conditions, and both are gates. The second is a matching rule: the money is paid only if the recipient chooses to take the project, which imports that recipient’s admission criteria into a municipal subsidy. The first is a jurisdictional one: the donor is reimbursed only for donating to an institution that the Chinese state recognises as national-level. Note what follows economically, and it is not a scandal — it is the predictable consequence of any subsidy with a residency requirement. If a firm’s open-source strategy is to donate to CNCF, the Apache Software Foundation or the Linux Foundation — the three bodies that in fact confer the international standing the policy’s own preamble says it wants, “global influence” — the CNY 3 million is not available. The firm has the same project, the same donation and the same public good, and it is paid for one recipient and not the other. That is not corruption; it is a price differential, and price differentials steer. The observable prediction is arithmetic: a subsidy worth CNY 3 million per project will move marginal donations toward OpenAtom and away from international foundations — not because anyone ordered it, but because one is reimbursed and the other is not. Beijing’s own 《北京市开源生态体系建设实施方案(2026—2028年)》 states the aim of cultivating ten internationally influential projects, of which no fewer than five are internationally leading — and this is the instrument by which a district intends to deliver them.
Fourth, the “national AI open-source community” clause, read as a governance claim rather than a funding claim. The policy undertakes to build a national AI open-source community, support models and toolchains to move in (入驻), and build an operations platform — with the park itself designated 自主可控 in software and hardware. “Move in” is a real-estate and tenancy verb, not an open-source one. A community that one joins by contributing to a repository has no tenant; a community that one moves into has a landlord, a lease and a set of house rules the landlord writes. Set this against the two facts already established in this series: the MIIT “AI + Software” action plan routes open-source infrastructure through the OpenAtom Foundation and the national AI open-source community AtomGit (confirmed by MIIT’s own published Q&A, 2026-09-11, which also commits to “strengthen the application of self-developed open-source licence agreements and compliance guidance”); and the 2026-11-14 briefing’s finding that HECOS 1.0 in Pakistan is the first case in this run where Chinese upstream generated a foreign downstream with its own stake in the upstream’s health. A community with a landlord and a sovereignty requirement is a distribution platform with exceptional reach. It is not the same institution as a repository whose maintainer list a stranger can join — and no clause in these ten actions pays for that list.
Fifth, the falsifiable test, and it is cheap to run. If, within the first subsidy cycle, the donor of a project reimbursed under action five retains or is granted maintainer rights in the receiving organisation’s governance — a seat, a vote, a nomination right — then the subsidy has bought a commons institution and not only a transfer. If the received projects enter a sovereign platform whose contribution process, maintainer roster and licence-approval path are controlled by the platform and the receiving body, then the state has bought an asset relocation with a public-good name, which is precisely what the 2026-10-26 entry in this series found to be the discriminating variable in donated corporate projects: the ones that lived were the ones where governance genuinely moved. The migration test is also available and cheaper than any of this: do the models and toolchains that “move in” keep a public mirror, a public issue tracker and an upstream-flowing patch path — or does moving in mean moving off?
Checkable questions.
(i) Which projects have actually been paid the CNY 3 million donation subsidy, to which receiving organisation, and what did the donation agreement say about governance? A donation that transfers copyright and nothing else is an asset sale with a tax advantage. The agreement text is the single most informative document in the entire policy and it is not published.
(ii) Is there any line in the 2026-2029 schedule — any of the ten actions — that pays an individual maintainer, reviewer or security responder, directly or through a fellowship? If the answer is no across an entire industrial policy, then the ecosystem’s funding architecture has been formally confirmed to be person-blind, and every “who answers the stranger” question this series has asked becomes a question about the tax code rather than about culture.
(iii) Does the “national AI open-source community” have a published contribution guide, a maintainer list and an OSI-approved licence before or after the models move in? Order matters here exactly as it did with the industrial-operating-system standards working group on 2026-11-15: tenancy first and governance second produces a platform; governance first and tenancy second produces a commons.
Sources: 北京经济技术开发区管理委员会 — 印发《北京经济技术开发区关于加快建设开源开放信创产业高地的若干措施》的通知(京技管发〔2026〕16号)(2026-11-02, published 2026-11-05, effective to 2029-12-31, superseding No. 3 of 2025; the ten actions; the 2027 targets of 1,000 xìnchuàng enterprises, CNY 100 billion revenue, 50 flagship solutions and no fewer than 50 high-quality open-source projects; action five’s CNY 1m project award, CNY 3m donation award, CNY 1m commercialisation award and CNY 3m event support; the national AI open-source community and its operations platform; the CNY 3m-per-year platform support; the 1:1 research matching capped at CNY 30m) · 北京经济技术开发区信息技术产业局 — 政策解读 (2026-11-05; the revision list of clauses one, three, four, five, seven and eight; the scope of the 225 km² Yizhuang New Town; the declaration portal) · 北京经济技术开发区 — 政策文件原件 (the 2026-11-02 drafting date and 2026-11-05 publication date; the same action-five text including the software- and hardware-sovereign phrasing of the national AI community) · 工业和信息化部信息技术发展司 — 七问+一图,读懂《“人工智能+软件"专项行动实施方案》 (2026-09-11; the routing of open-source infrastructure through the OpenAtom Foundation and the national AI open-source community AtomGit; first-release in the community; the open-source zone for intelligent operating systems; application of self-developed open-source licence agreements and compliance guidance; open-source contribution into research evaluation and talent recognition) · This series’ 2026-11-15 briefing — the committee arrives before the community · This series’ 2026-11-14 briefing — the first export that is a distribution · This series’ 2026-11-11 briefing — the instrument that fits the transaction
⚖️ Legal & Licensing — The Behaviour Condition, the Scale Condition and the Subject Condition
1. The document. On 2026-11-14 the platform account 观察 published a long analysis, 《“乌镇共识"的背后:AI开源,正在从"无国界"变成"有条件”》, on what happened to open-weight licensing during 2026. It is not an official document, and this entry treats it as an industry reading rather than a policy source — but it is the most complete Chinese-language survey of the licence texts this briefing has found, it quotes the licences rather than the marketing, and its central distinction is correct and portable.
2. The distinction, which is the entry. The piece opens by dismantling a premise: the claim that software licensing was ever “borderless, unconditional and irrevocable” was always idealised. GPL is the strongest possible condition — modify and you must open; Apache 2.0 carries a patent-retaliation clause that terminates the patent grant on suit; MIT requires the notice be kept; and open source has been subject to export control for years (cryptography; GitHub’s sanctions-based access restrictions after 2019). The real difference is not whether conditions exist but what they attach to. Software licences attach conditions to behaviour — did you modify it, how did you distribute it, did you attribute, did you sue. They do not attach to identity: a Russian developer, an American company and an African startup that behave identically receive identical rights. AI open-weight licences have added two further layers — a condition on scale, and a condition on subject.
3. The clause inventory, as the piece assembles it. Scale: Meta’s Muse Spark Open (released 2026-10-15) is free to download, modify and run locally, with three exemptions — internal use, research use, and commercial activity that does not serve the model over an API — but once the company and its affiliates exceed USD 20 million in any rolling twelve months and serve the model via API, it must notify Meta and sign a separate written agreement, and the product UI must carry “Built with Muse Spark Open”. Earlier Llama generations set a 700 million monthly-active-user commercial threshold. Subject: MiniMax’s M2.7 video-model licence excluded the United States, the European Union, the United Kingdom and South Korea, the company explaining the US carve-out as a response to Hollywood copyright litigation and the rest as regulatory uncertainty, with excluded parties able to apply separately and undertake compliance. Both at once, among Chinese vendors: Kimi K3 requires a separate agreement where the company and affiliates exceed USD 20 million in trailing twelve-month revenue and operate MaaS; Qwen3.8-Max raises the threshold to USD 50 million, extends the restriction to a separately defined “AI Work Assistant” category (coding and office-productivity tools), counts the licensee plus affiliates, and pairs a flagship-with-a-gate and a mid-size-with-Apache-2.0 as an explicit two-track strategy; Qwen3.8-Flash-Next switched to the Qwen Community License 1.0, abolishing the revenue threshold and requiring separate authorisation for MaaS or AI Work Assistant use whatsoever; MiniMax M3 tiered it — below USD 20 million of related product revenue, label and notify; above, negotiate — measuring the threshold on the product line rather than the company; and DeepSeek V4-Pro and the GLM-5.3-Flash line stayed on MIT with no revenue or use restriction.
4. The design logic the piece names, and it is the sharpest sentence in it. “成功惩罚条款” — a success penalty clause. The threshold is calibrated to have no effect whatever on a startup in its early stage and to detonate exactly at the moment the business works. The three exemptions in Muse Spark Open are not generosity; they are actuarial: they leave the non-paying part of the market free and set the trigger at the instant the money starts. And the piece documents the same architecture on the Chinese side with more precision than any official statement: the separate agreement for Kimi K3 reportedly carries revenue-share terms reported up to 30%; Reuters-reported negotiations with Microsoft, Amazon and Google over K3 on Azure, AWS and Google Cloud were aimed at up to 30% of related service revenue. The observed consequence is stated bluntly: the terms do not currently bind startups at all. They bind Volcano Engine, Alibaba Cloud and Baidu AI Cloud — and the licence’s present function is to constitute a negotiation mechanism, on the Unreal Engine model.
5. The framing, from Wuzhen, and the honest part. The piece is careful about what the Wuzhen initiative is: the eighteen-point 《携手构建全球开放人工智能行动倡议》 was issued by the WIC’s AI committee, so it is a framework led and proposed by the Chinese side, not a negotiated multilateral text — “Wuzhen consensus” describes how it was used in the room, not what it is. Its operative sentence on openness is the one that matters: “strengthen the traceability of technology, and make model parameters, training-data provenance and training process more transparent, so that every open-source model has a clear origin and an auditable destination (来路清楚、去向可查)”. Wang Jian (王坚) of Zhejiang Lab, chairing the AI Wuzhen Dialogue, put the same thing in one line: open source pursues openness, but the precondition of openness is that it be trustworthy, controllable and governable (可信、可控、可治理). The piece’s own conclusion is the exchange-rate question this series would ask: whoever defines “trustworthy” sets the price of “open,” whether that definer is Brussels, Washington or Beijing — and the more specific the list, the more concentrated the discretion.
Institutional Analysis & Commentary:
Yesterday’s briefing ended by asking who is entitled to say no to a change and who pays them. Today’s legal entry supplies the missing half of that question from the other side of the transaction: in 2026, the licence no longer grades the act, it grades the actor — and a commons institution is precisely the thing that refuses to do that.
First, state the shift in property terms, because that is what it is. A copyright licence is a partial transfer of a bundle of rights: it alienates use, modification and distribution under conditions the buyer can verify by inspecting their own conduct. That is why it scales to strangers — conduct is self-observable, so a licence that conditions on conduct requires no registrar. Conditioning on revenue requires an accountant; conditioning on identity requires a registrar; conditioning on jurisdiction requires a border. Each new condition moves the licence one step closer to the world of Williamson’s asset-specific bilateral contract, where terms are set by negotiation between known parties rather than by published text between unknown ones. The Chinese piece notices this precisely and calls it what it is: from published standard text to one-to-one contract. **What it does not say, and should, is that one-to-one contracts do not have a commons property; they have a customer list.
Second, the “success penalty” is not a paradox — it is the price of an unsolved funding problem, and naming it correctly changes the policy conclusion. The critique most often made of these licences is hypocrisy. That is the wrong frame. A frontier model costs something on the order of a hundred million dollars of compute to train, cannot be recalled once its weights are published, and generates no per-copy revenue. A vendor that gives it away must recover the cost from the only observable that correlates with the value captured: the licensee’s revenue. The threshold is set where it is not because the vendor wants startups to fail but because a threshold set lower would tax the ecosystem’s reproduction and a threshold set higher would collect nothing. Compare this with the classical picture of a public good, which is financed by a tax base — a government that can charge everyone, including those who never use the good. What the success penalty reveals is that open weights have no tax base: there is no institution that can collect from the beneficiaries who do not sign anything. This is not a Chinese problem, and today’s evidence is that it is not a vendor problem either — Meta and Kimi converged on USD 20 million independently. The convergent number is the empirical signature of a missing institution, not of a shared conspiracy.
Third, the Chinese evidence contains one genuine counter-example and one genuine warning, and they must be kept apart. The counter-example: DeepSeek stayed on MIT from V3 through V4-Flash, and GLM-5.3-Flash likewise, in a half-year when every comparable frontier family added a gate — the industry piece’s own word for it is 逆势 (against the tide), and its explanation is the interesting part: either a deliberate strategy of trading maximal openness for ecosystem, or a cost structure different enough that the vendor does not need licence revenue. Both explanations are testable and neither is ideological. The warning is the parts of the inventory not usually quoted together: MiniMax M2.7 changed its licence six hours before release from a commercially friendly form to written authorisation required for all commercial use; Gemma 4 moved to Apache 2.0 in April 2026 while Gemma 3 and earlier kept the restrictive custom terms, with Google retaining a unilateral right to update prohibited-use policy and to require developers to enforce Google’s rules in all Gemma-derived projects; and the threshold metrics are not comparable across vendors at all — company revenue, licensee-plus-affiliate revenue, related-product revenue and monthly active users are four different denominators, so “under USD 20 million” means four different things. From the buyer’s side, the most consequential legal fact of 2026 is not that AI licences are conditional. It is that they are no longer legible without counsel, and the class of actor that can afford counsel is not the class that the open-source way was built to include.
Fourth, the sovereignty dimension, treated as an institutional fact rather than a grievance. The EU AI Act’s Article 53(2) exemption for free and open-source models applies only where the licence permits access, use, modification and distribution and the weights, architecture and usage information are published, and it does not apply to models with systemic risk. The US EAR currently excludes publicly available open weights from control, and the discussion under way concerns the upstream — advanced-chip controls and the extension of the foreign direct product rule. Both regimes therefore define “open source” not by the OSI’s four freedoms but by a list of conditions on the artefact and its provenance, and both can be revised without the licence changing at all. This is the structural insight the Chinese piece gets right and most Western commentary misses: the binding constraint on an open-weight licence is not the licence text, it is the customs regime that determines whether the download can be sustained. A vendor conditional and a regulator conditional are both setting terms; only one of them publishes them.
Fifth, the falsifiable prediction, and one that can be settled by reading a file rather than an essay. If the scale condition is genuinely a negotiation mechanism rather than a revenue product — as the piece argues from the fact that only large platforms can trigger it — then within twelve months at least one Kimi, Qwen or MiniMax-style threshold licence should be amended to reduce friction for small licensees (a published safe-harbour, a self-certification form, standard terms with pre-set percentages) as the vendor discovers the legal cost of an unclear gate. If instead the thresholds get more granular and the exceptions get narrower — a new defined category, a new affiliate definition, a new jurisdiction carve-out — then the trajectory is toward bespoke contracts, and the 2026-10-23 entry’s concern about “permissive washing” will have been the early and mild form of it. And the decisive marker, unchanged from 2026-11-13: whether any of these licences is submitted to an OSI or equivalent conformance process, or whether the word “open source” simply continues to be used in the press release while the text says “open weights.” On the piece’s own analysis, most of the inventory fails the OSAID condition of use for any purpose without separate permission, and the vendors themselves mostly say “open weights” when they are being careful.
Checkable questions.
(i) Which of these thresholds is company revenue, which is licensee-plus-affiliate revenue, which is related-product revenue, and which is monthly actives — and does any vendor publish a worked example of who is and is not in scope? Four denominators with the same numeric gate is not a threshold; it is four thresholds, and a startup cannot budget against a range.
(ii) Has any Chinese vendor’s threshold licence been amended after release, in which direction, and with how much notice? MiniMax’s six-hour pre-release reversal and M3’s subsequent relaxation are the only two data points available. A pattern of post-release tightening would mean the published text is not the contract.
(iii) Does the eighteen-point Wuzhen initiative’s traceability language — parameters, data provenance, training process, “clear origin and auditable destination” — enter any actual licence, procurement specification or conformance test within twelve months? A principle that stays in a communiqué costs nothing. A principle that enters a procurement annex sets the price of admission for every open-weight model in the world.
Sources: 观察 — 观察|“乌镇共识"的背后:AI开源,正在从"无国界"变成"有条件” (2026-11-14; Wang Jian’s Open Artificial Intelligence translator note and his formal statement that openness presupposes trustworthiness, controllability and governability; the eighteen-point WIC AI committee initiative and the “clear origin, auditable destination” formulation; the GPL/Apache-2.0/MIT condition analysis and the point that conditions attach to behaviour not identity; Meta Muse Spark Open’s 2026-10-15 release and its USD 20m-plus-API trigger and attribution requirement; Llama’s 700m MAU threshold; Gemma 4’s April 2026 Apache 2.0 move with Gemma 3 terms retained and Google’s unilateral policy-update right; Kimi K3’s USD 20m MaaS threshold and reported revenue shares up to 30%; Qwen3.8-Max’s USD 50m threshold and the AI Work Assistant category; Qwen3.8-Flash-Next and Qwen Community License 1.0; MiniMax M2.7, H3 and M3’s licence trajectory; DeepSeek V4 and GLM-5.3-Flash on MIT; the “success penalty clause” reading; EU AI Act Article 53(2); the EAR treatment of publicly available weights and the foreign direct product rule discussion; Qwen’s 460-plus models and 4bn downloads) · DoNews — 开源大模型商业化转向:许可证设营收门槛与MaaS限制 (2026-08-27; the parallel inventory with the Qwen Community License 1.0 switch, the MiniMax H3 jurisdiction exclusions, GLM-5.3-Flash’s MIT licence and its 2025 revenue split of CNY 534m localised deployment against CNY 190m cloud API, the reported Microsoft/Amazon/Google revenue-share talks at up to 30%, and the finding that the clauses currently bind large platform firms rather than startups) · 新华社 via 凤凰网 — 综述|中国开源模型在全球产业端加速落地 (2026-11-12; Web Summit Lisbon; Revolut replacing OpenAI and Anthropic coding assistants with Zhipu’s GLM for roughly 2,000 engineers, about 60% less coding time and near-90% lower AI call cost; Orange, Renault and Siemens publicly using DeepSeek and Qwen as supplements; the deployment-on-own-servers motive) · 新华社 — 从这场乌镇对话,读懂构建AI开源生态关键词 (2026-11-04; Wang Jian’s beneficiary-to-contributor formulation in the AI Wuzhen Dialogue) · 通信世界 via 腾讯新闻 — 王坚:人工智能时代中国成为开源重要贡献者 (2026-11-10)
🏗️ Institutional Change — The Instrument That Was Already in the Table
1. The document, and the round trip it reveals. MIIT’s “AI + Software” special action implementation plan (《“人工智能+软件"专项行动实施方案》, 工信部信发〔2026〕209号), read through the ministry’s own published Q&A on 2026-09-11, devotes its fifth question to open source. Three measures. (a) Build out open-source infrastructure: route it through the OpenAtom Foundation and the national AI open-source community (AtomGit), guide firms to open up intelligent programming, software-product intelligence and agent outputs, support key projects to make their first release (首发) in the community, establish open-source zones for intelligent operating systems and other fields, and strengthen the application of self-developed open-source licence agreements and compliance guidance. (b) Establish a selection and cultivation mechanism for high-quality projects, encouraging diversified business models and deep application of good projects in manufacturing and other key sectors. (c) Stimulate participation: support universities, research institutes and enterprises in putting open-source contribution into research evaluation and talent recognition, and explore using open-source contribution as a reference indicator for identifying high-quality SMEs in software.
2. Why this is an institutional event rather than a funding line. The third measure is the one that matters, and its significance is that it is the same instrument the 2026-11-02 Beijing subsidy is trying to buy with cash, and it was already on the table two months earlier — in a ministry document, addressed to the evaluation system rather than to the bank account. “Open-source contribution into research evaluation and talent recognition (科研评价与人才认定体系)” is a change to what counts; the BDA schedule is a change to what pays. A contribution that counts for a promotion or a professorship is worth more to an academic than CNY 1 million to an institution, and it costs the state nothing. This is the first time in this series that a Chinese policy document has named the academic incentive system as the lever for contribution — and it is precisely the lever that the 2026-11-14 entry on HarmonyOS found being pulled in the opposite direction, where a national competition ranking made using a platform count for students while no course required a merged patch to pass.
3. The Chongqing precedent, for the record, because it shows the same measure arriving through a different door. The OpenAtom Foundation’s own policy digest records that on 2025-12-13 the Chongqing municipal government issued its “AI+” action plan, which states: “encourage enterprises to join with developer communities and open-source foundations to build open-source communities, open deep-learning frameworks and pre-trained models … push universities to explore including teachers’ contributions made in open-source communities within the reference indicators of professional-title assessment and performance appraisal (职称评定、绩效考核).” The same digest, which covers roughly thirty provincial and municipal instruments in a two-month window, shows the Chinese open-source policy stack in its actual shape: Fujian encouraging AI open-source communities and public service platforms; Guangdong supporting OpenHarmony ecosystem building and joint university–enterprise open-source centres; Anhui paying up to CNY 5 million in equipment subsidies for, among other platforms, an open-source ecosystem centre (社区); Guangxi subsidising embodied-AI datasets with a 10% purchase subsidy capped at CNY 1 million; Henan and Tianjin supporting the establishment of open-source promotion institutions (开源促进机构) inside free-trade zones, in the course of aligning with high-standard international economic and trade rules; Shanghai’s December 2025 open-source system-building plan with its 2027 targets of one to two internationally influential communities, 100 open-source commercialisation firms, 200-plus quality projects and 3 million developers, plus a pilot of contribution-oriented talent evaluation (以开源贡献为导向的人才评价机制); Beijing municipality’s 2026-2028 plan with ten internationally influential projects of which at least five internationally leading; and Changsha paying up to CNY 1 million for a qualifying shared open-source project.
4. The Shanghai implementation, reported a week ago. On 2026-11-11 Economic Information Daily (Xinhua) reported Shanghai’s drive to be “the city that understands developers best (最懂开发者的城市)”, built on its December 2025 plan: up to CNY 20 million for foundational development-tool R&D, up to CNY 5 million for a high-quality open-source project, and the contribution-oriented talent-evaluation pilot. The concrete move: on 2026-10-30 OSChina (开源中国), the largest domestic open-source platform company, moved its headquarters to Zhangjiang in Pudong, and launched a “Model Ark Base (模力方舟基地)” offering community operations, platform-resource matching, scenario validation and incubation. OSChina’s CEO Xu Yong (徐勇) gave the reason in the paper’s own words: Shanghai has the two scarce inputs — dense real industrial scenarios and systematic institutional supply — and Pudong is the country’s first AI innovation-and-application pilot zone. The company committed to three things: bring eighteen years of developer-operations experience to Shanghai, open the Model Ark model interfaces and compute resources to Shanghai firms and developers, and run continuous AI-tool training. Its 2026 Shanghai Open-Source Software Application Innovation Competition, launched 2026-08-05, drew 766 registrations, 9 organisations posting 12 real industry problems, 269 projects attacking them, 10 teams per track in the finals, one third of them Shanghai teams, with tracks aligned to the city’s integrated circuits, high-end equipment, intelligent vehicles and intelligent compute.
Institutional Analysis & Commentary:
This is the week’s most important document and it is the least dramatic one: a ministry Q&A, published in September, saying that open-source contribution should count in academic evaluation. Read against the BDA subsidy schedule and the Hangzhou talent standard, it locates the actual lever of Chinese open-source policy — and it is not money.
First, name the mechanism precisely, because it is the difference between a policy that works and a policy that is expensive. A subsidy is a price on an act, and prices on acts have to be paid every time the act recurs. An evaluation criterion is a price on a status, and it is paid once, in a document, and then collects itself: every academic in the country recalculates their portfolio against it without a single further transfer. If a merged patch counts toward a professorship, the supply of merged patches rises at zero marginal fiscal cost — and the same is true of the BDA donation subsidy at a marginal cost of CNY 3 million per project. This is why the Chinese ecosystem’s instrument choice has been so consistently skewed toward certification, curriculum, standards and platforms: those are the instruments that write evaluation criteria. The certification examined on 2026-11-13 is a private attempt to create one; the HarmonyOS competition’s entry into the national discipline ranking, on 2026-11-14, is a state one; the Hangzhou OPC nine-level talent standard, on 2026-11-15, is a municipal one. MIIT’s Q&A is the first that points the same lever at the thing that actually needs supplying.
Second, and immediately, the hard question: contribution to what? The clause says open-source contribution without specifying a repository, a review standard, or a verifier. Every evaluation criterion requires a metric, every metric requires an instrument, and every instrument in this ecosystem so far has counted activity that a body can see: certificates issued, courses delivered, projects hosted, events held. The failure mode is not hypothetical and it is not Chinese. Faculty evaluation systems worldwide that adopted “contribution” metrics produced commit counts, star counts and conference talks, because those are countable and merged-and-reviewed patches are not — a patch’s quality is a judgement, and judgements require a reviewer with standing, which is exactly the institution in short supply. So the testable question is not whether universities will adopt the criterion — MIIT and Chongqing and Shanghai have all now said they will. It is whether the verifier they adopt can distinguish a merged patch from a repository that a student was told to open. The three markers from the 2026-10-26 entry on donated projects apply here directly: is there a CODEOWNERS file with an outside name, an accepted proposal not authored by the original team, a production adopter in the documentation. Those are countable, hard to fake, and already published.
Third, the Shanghai case is the cleanest available specimen of the ecosystem’s own theory of itself, and it deserves a fair reading. “Dense real industrial scenarios plus systematic institutional supply” is, as a diagnosis, correct — and it is a diagnosis of why a platform city works. OSChina is not a foundation and not a vendor of a stack; it is a distribution layer: nineteen years of developer traffic, model interfaces and heterogeneous domestic compute (Ascend, Muxi, Tianshu) adapted behind one API. Read the economic function carefully and it is the same function that the standards working group on 2026-11-15 and the “move-in” community on 2026-11-02 perform: reduce the friction of the transaction that already has two parties. Its competition design is unusually good and worth crediting: nine organisations posted twelve real production problems, not the platform’s problems — that is demand-side specification, which is closer to how a commons actually forms than any number of code camps. But note what the platform is and is not: Model Ark makes the base layer light so developers can go deep above it, and its incentive is to maximise the number of models, users and scenarios flowing through one distribution point. A distribution layer is a superb institution. It is also the one institution whose rational strategy is to keep the contribution relationship on its own side of the interface. Nobody in that arrangement has a reason to hand a stranger merge rights into the base layer; they have a reason to hand them a very good API.
Fourth, the comparative point that the week makes possible. Three documents, one fortnight, three different prices for the same act. MIIT (September): contribution counts in research evaluation — price zero, instrument an academic career. Shanghai (December 2025, implemented in November 2026): a quality project is worth up to CNY 5 million and a foundational tool up to CNY 20 million, with a contribution-oriented evaluation pilot — price high, instrument a grant. BDA (November 2026): a donation to a national-level organisation is worth CNY 3 million, a project up to CNY 1 million, hosting a national event up to CNY 3 million — price explicit, instrument a disbursement. And Chongqing (December 2025): a teacher’s community contribution counts toward their title — price zero, instrument a personnel file. The variance is not confusion; it is division of labour across levels of government, and it is unusually legible. What it is not, in any of the four, is a price on the specific act of reviewing a stranger’s patch, because that act has no payee — it is performed by a person inside a project, and none of these instruments has a line that reaches a person inside a project except through an employment or an academic relation.
Fifth, the prediction, stated so that it can fail in a way that matters. If Chongqing’s and Shanghai’s contribution-into-evaluation pilots produce, within two academic years, publicly listed promotion cases cited on a project’s own governance record — a named maintainer whose institutional promotion cites commits she merged upstream — then the ecosystem will have found its cheapest and most durable commons institution, and the “who answers the stranger” problem becomes a personnel-policy problem with a known solution. If instead the pilots produce contribution counts aggregated by platform (commits, projects, downloads), the metric will have been captured by the party that owns the counter — and the party that owns the counter is, in this week’s evidence, increasingly a city, a platform company and a foundation, none of which needs merge rights to score the number.
Checkable questions.
(i) Which specific universities have issued implementation rules under MIIT’s and Chongqing’s contribution-recognition language, and what exactly do those rules count? A rule that counts merged patches to named repositories is a commons instrument. A rule that counts participation in open-source activities is an attendance register with a title attached.
(ii) Is there a published registry of the projects selected under any “high-quality open-source project” scheme — Beijing, Shanghai, Changsha, Anhui — with the selection criteria and the reviewers’ names? Four municipalities are now paying for the same designation. Without a published criteria document, “high-quality” is a discretionary term, and discretionary terms in subsidy programmes select for firms with the best applications rather than the best projects.
(iii) Does OSChina’s Model Ark Base, or any grant-funded distribution layer, publish an upstream contribution path — can a Shanghai developer’s fix reach the model or toolchain vendor through it, or does the platform terminate the relationship? A distribution layer that terminates the relationship is a shop. One that passes patches back is a downstream, which is the institution that HECOS 1.0 accidentally created in Pakistan and that no procurement contract in this series has yet produced.
Sources: 工业和信息化部信息技术发展司 — 七问+一图,读懂《“人工智能+软件"专项行动实施方案》(2026-09-11; 工信部信发〔2026〕209号; the open-source measures routing through OpenAtom and AtomGit, first-release support, intelligent-operating-system open-source zones, self-developed licence application and compliance guidance, project selection and cultivation, and open-source contribution into research evaluation and talent recognition plus as an SME-identification reference) · 经济参考报 / 新华网 — 上海推进开源平台建设 打造"最懂开发者的城市” (2026-11-11; the December 2025 Shanghai open-source system-building plan and its 2027 targets; up to CNY 20m for foundational development tools and CNY 5m for a quality project; the contribution-oriented talent-evaluation pilot; OSChina’s 2026-10-30 headquarters move to Zhangjiang and the Model Ark Base; Xu Yong’s scenario-density and institutional-supply argument; the three commitments; the 2026 Shanghai competition’s 766 registrations, 12 industry problems from 9 organisations, 269 responding projects and one-third Shanghai finalist share) · 开放原子开源基金会 — 政策观察|开源及人工智能政策概览(11-12月)(the provincial and municipal instrument digest: MIIT’s AI+Manufacturing action plan text on open-source licences and “new open-source rule order”; Fujian, Guangdong, Hubei, Anhui with its up-to-CNY-5m open-source ecosystem centre subsidy, Chongqing’s AI+ plan with the teacher-contribution clause, Guangxi, Henan, Shanghai, Hangzhou, Chongqing’s AI-terminals plan, Yunnan, Changsha’s up-to-CNY-1m shared open-source project award, Beijing’s 2026-2028 open-source ecosystem plan with ten internationally influential and five internationally leading projects, and Tianjin’s international open-source promotion institution inside the FTZ) · This series’ 2026-11-13 briefing — the certification fee and the missing maintainer rung · This series’ 2026-11-15 briefing — the committee arrives before the community
🔍 Commentary — Who Signs the Permission Slip
First, the week’s three documents are one document about who is allowed to be a counterparty. Beijing Yizhuang will pay CNY 3 million for a project donated to a national-level open-source organisation and accepted. A Kimi or Qwen licence will grant rights freely until the licensee crosses USD 20 or 50 million and then require a signature. MIIT will count a contribution toward a professorship, and Shanghai will count it toward a talent designation, and Hangzhou will grade a person on a nine-level OPC capability standard. Every one of these instruments answers the same question — who is the counterparty, and what documentary evidence establishes that they are one. None of them answers the question that a commons asks, which is who may act without asking permission.
Second, this is why the “no budget line for the inbound half” diagnosis was right but incomplete, and the correction matters. The inbound half — reviewing, triaging, answering, fixing at three in the morning — is not unfunded because the ecosystem is stingy. It is unfunded because it is not a transaction that any of these instruments can express. A grant needs a grantee, a subsidy needs an applicant, a tax break needs a filing, an evaluation needs a dossier. A reviewer who reads a stranger’s patch and says no, and here is why produces a decision that belongs to no legal person, generates no invoice, and — critically — is not even measurable in a way that survives audit, because the value of the reply is in the recipient’s next attempt, not in the reply. Everything an administrative state can see is a party. A commons is made of acts by people who are not parties to anything. That is the whole of the structural finding, and this week it is visible in a municipal subsidy schedule, a licence threshold and an academic evaluation clause simultaneously.
Third, and this is the part that should change the reader’s expectation about where relief will come from. The most commons-shaped instrument in today’s material is not the biggest one. It is Chongqing’s sentence about teachers’ community contributions counting toward professional-title assessment, and the reason is not generosity but structure: an academic is a person with a long horizon, a named identity and an employer that will survive the next funding cycle. A person can be a maintainer. A project cannot. A platform will not. The entire apparatus examined this fortnight — foundations, committees, standards, certifications, blue books, typical cities, nine-level talent ladders, donation subsidies — is composed of organisations, and organisations are structurally incapable of being the thing that answers a stranger, because a stranger’s patch is addressed to a name. The moment a policy document makes a name’s career depend on merged-and-reviewed work rather than delivered-and-counted work, it has accidentally built the missing institution, at no fiscal cost.
Fourth, the honest caveat, and it is serious enough to state against the thesis. Metrics of contribution, once they enter a promotion file, will be gamed in exactly the ways that metrics always are — this series has already recorded the mechanism (star-to-follower ratios above 200:1 as the signature of a launch-driven repository; certificates as purchasable credentials; courses that end in an app rather than a patch). There is a real risk that “open-source contribution” becomes, within two years, a résumé line produced by a consultancy, and that the fix will be harder than the problem. The only available defence is the one this series has been recommending since 2026-10-26 and it costs nothing: verify against the repository, not the certificate. A CODEOWNERS entry with a name outside the founding company, an accepted proposal not authored by the original team, a merged pull request with a review thread. Those three are public, they are already published for every serious project, and they cannot be manufactured by the applicant alone — because manufacturing them requires someone with merge rights to say yes, which is the act the whole system has been unable to buy.
Fifth, the sentence to keep, and it came from Hangzhou a day before the last briefing was written. KAIYUANSHE’s own call asked “how does a code of conduct actually get enforced rather than published?” Three documents later, the answer has a shape: enforcement requires a person whose judgement is final and whose position does not depend on the person they are judging. Yizhuang’s schedule pays the donor, Qwen’s licence prices the licensee, MIIT’s Q&A credits the contributor — and each of the three, for the most respectable of reasons, keeps the final judgement in the organisation. The next instrument this ecosystem builds will not be a bigger subsidy or a stricter licence. It will be — if it comes — a personnel rule that lets a named individual in a Chinese institution spend professional hours reviewing code they did not write, for people they will never meet, and be promoted for it. That is the whole missing piece, it costs nothing, and no document in this fortnight has proposed it except in the one place nobody was watching: a line about teachers in a municipal AI plan.
Sources: 北京经济技术开发区 — 开源开放信创产业高地若干措施 (2026-11-02; the CNY 3m donation award, the CNY 1m project and commercialisation awards and the national AI open-source community) · 观察 — “乌镇共识"的背后:AI开源,正在从"无国界"变成"有条件” (2026-11-14; the success-penalty clause and the behaviour/scale/subject distinction) · 经济参考报 — 上海打造"最懂开发者的城市” (2026-11-11; the contribution-oriented talent-evaluation pilot and OSChina’s distribution layer) · 开放原子开源基金会 — 开源及人工智能政策概览(11-12月)(the Chongqing AI+ plan’s teacher-contribution clause and the thirty-instrument provincial stack) · 工业和信息化部 — “人工智能+软件"专项行动实施方案解读 (2026-09-11; open-source contribution into research evaluation and talent recognition) · 开源社 — COSCon'26 call for participation (“How does a code of conduct actually get enforced rather than published?”) · This series’ 2026-10-26 briefing — what survives a donor · This series’ 2026-11-14 briefing — eight hundred courses and no governance textbook