⚠️ 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 — 2026-10-31

🏛️ Policy & Regulation — The Pilot-Test Base: What China Built to Answer the Question a Progress Report Left Open

1. The question this file answers. Yesterday’s briefing ended on a sentence from the Progress Report on the AI Capacity Building Inclusive Plan (2026-10-24, Ministry of Foreign Affairs domain, Global AI Innovation Governance Centre): open cooperation requires a supporting knowledge framework, “including toolchains, test environments and related support that remain to be built” (亟待建设的工具链、测试环境和相关支持). The checkable question posed then was whether the 2027 report would contain an entry that is a repository rather than a platform. That question stands. But the Chinese-language reporting of the past month, read in sequence, describes an institutional form built to answer the test-environment half of it — and the form is not a repository. It is called 国家人工智能应用中试基地, the National AI Application Pilot-Test Base.

2. What the form is, in its own official language. The base is not a 2026 invention. It has a policy parent: the State Council Opinions on Deepening the Implementation of the “AI+” Initiative (《国务院关于深入实施"人工智能+“行动的意见》, 2025-08-26), which called for open-source ecosystem flourishing and for open-source contributions to be recognised in university credit and faculty appraisal. Its first sectoral deployment came the same autumn, on 2025-10-20, when five ministries and commissions — the National Health Commission, NDRC, MIIT, the National Administration of Traditional Chinese Medicine and the National Disease Control Administration — issued joint implementation opinions on “AI+ healthcare” naming the base explicitly: “结合国家人工智能应用中试基地，支持省级统筹建立行业公共支撑服务平台，提供统一、高效、开放的人工智能算力服务” — in conjunction with the national AI application pilot-test base, provinces shall establish sectoral public support service platforms offering unified, efficient and open AI compute. The same document adds: “推动医学人工智能开源软件建设，持续提升共建共享水平” — promote the building of medical AI open-source software and continuously raise the level of co-construction and sharing. Then, in September 2026, MIIT and eight other departments issued the 15th Five-Year Plan for the intelligent connected new-energy vehicle industry, one clause of which reads: “加快汽车人工智能应用中试基地建设” — accelerate the construction of the automotive AI application pilot-test base. That clause has now been executed.

3. Base one — automotive, published 2026-10-23. At the 2026 World Intelligent Connected Vehicle Conference on 2026-10-23, China Mobile, as lead builder, published the industry’s first automotive AI foundation (汽车行业AI基座) on behalf of the base, witnessed by the Beijing Economic-Technological Development Area administrative committee, Tsinghua University, FAW, Dongfeng, Changan, BAIC and Li Auto — more than ten organisations. The stated position is an institutional one, not a technical one: “中立不造车、能力全开放” — neutral, does not build cars, capabilities fully open. Three capability layers: compute (112.7 EFLOPS of managed capacity, callable on demand), data (OEM core data ownership fully retained, flowing only under a compliance framework) and model (a full toolchain from development to validation to deployment). Six agents cover R&D, manufacturing, intelligent driving and testing. The reported deployments: a design cycle 3× faster; a process card that took 5 days now produced in 4 hours; visual inspection defect detection above 99.5% with line adaptation cut from 4 weeks to 1 week; 20-plus fault classes recognised and downtime cut 30%; driving-policy training accelerated 2×. All described as already landed at more than ten central- and state-owned automakers.

4. Base two — household appliances, launched 2026-10-27. In Qingdao on 2026-10-27 the National AI Application Pilot-Test Base (Household Appliances direction) held its construction launch with 200-plus experts from appliance firms, AI firms and universities. Its goal: “打造全球引领的’新一代智慧家电’中试基地” — a globally leading pilot-test base for next-generation smart appliances — moving the industry from “单打独斗” (fighting alone) to “联合攻坚” (joint assault). Five capabilities: product pilot-testing, scenario pilot-testing, industry pilot-testing, scaled promotion, standards and safety. The named host is the decisive detail: “作为基地承建方，海尔智家将依托产业实践与场景资源，联合行业各方推进基地建设和运营” — Haier Smart Home, as the construction undertaker (承建方), will rely on its industrial practice and scenario resources to advance construction and operation. First-batch ecosystem partners signed on stage: Huawei, China Mobile, Alibaba Cloud. Experts appointed from the China Household Electrical Appliances Association, CAICT, Xi’an Jiaotong University and iFlytek. The operational mechanism is the most institutionally interesting line in the announcement: a supply-demand operations platform collects firms’ real requirements and distils common problems, then organises universities and institutes to attack them jointly “以’揭榜挂帅’等方式” — by open-call contract; validated solutions then go onto an application service market. The summarising formula: “需求从行业中来、成果到行业中去” — demands come from the industry, results go back to the industry.

5. Base three — the one with a public inventory, opened 2026-09-27. The National AI Application Pilot-Test Base (Culture and Tourism) campus opened in Hangzhou’s Gongshu district on 2026-09-27, described as the country’s first and only national base in that direction: 142,800 m², architecture 1+4+5+N (one integrated portal; four domains — cultural preservation, production, dissemination, tourism integration; five function centres — compute, data, model, application, scenario validation). Its opening published an unusually specific inventory. Five models: the Wenrun culture model trained on museum archives, heritage records and classical texts; a Wenrun·Yuanzheng ancient-building inspection model recognising 10-plus defect classes at above 87% accuracy; three specialist models. Four datasets: an ancient-building gene dataset aggregating data on 40,000-plus structures nationwide and 400,000-plus named components — doors, windows, wood carvings; a public cultural space dataset covering 100-plus museums and memorial halls across 22 provinces; a tourism and audiovisual dataset; a cultural-outbound dataset. Five application scenarios: the Aiyouyou tourism super-agent, a copyright-and-data protection pilot platform, a one-stop inbound-tourism service, an AIGC creation platform, a full-media production system. Tenancy: 100-plus firms negotiating, 24 registered, 18 in office. Three-year targets: 15PB+ of sector data, 33 leading-edge scenarios, 15 industry standards. The district added a subsidy schedule — R&D support up to RMB 30 million; talent-project support up to RMB 15 million plus RMB 15 million of loan interest subsidy and RMB 5 million of equity financing — and two instruments named precisely: compute vouchers and token vouchers (算力券 + Token券).

6. The other three, for completeness of the count. The medical base (Guangdong/Guangzhou), landed 2025-07, reported on 2026-10-22 a matrix of 84 vertical models across clinical, public-health and public-service use; 154 high-quality datasets and 381 AI medical scenarios assembled from the province’s 68,000-plus health institutions and 182 tertiary-A hospitals; a shared disease-control compute cloud and trusted data space with 32 provinces in the pipeline and 23 having deployed connectors; and a lung-cancer pathology-genomics model cutting a targeted-gene test that cost tens of thousands of yuan to a few hundred, in as fast as one minute. The employment base held a seminar in Fuzhou on 2026-10-22/24 with Huawei, iFlytek and Shanghai AI Lab on corpus construction, sector model training and — notably — evaluation: “优先搭建评测标准与测试集，常态化开展回归评测” — build evaluation standards and test sets first, and run regression evaluation as routine. And the culture and tourism base above is the third with a public inventory.

Institutional Analysis & Commentary:

Begin with the thing that is easy to miss because it is boring, and it is the most important sentence in this file.

Every one of these announcements states a host. Not a site, not a funding line — a named organisation carrying the legal and operational weight. The medical base in Guangdong. The automotive base at China Mobile. The household-appliance base at Haier Smart Home, in the explicit role of 承建方. For culture and tourism, a district government with Zhejiang Provincial Cultural Investment Group. This is not decorative attribution. In Williamson’s terms it is the answer to the question L3 governance must answer and L4 allocation never does: who bears the residual. A test environment is not a document; it is standing capacity — clean rooms, calibration rigs, regression suites, connectors into other people’s data, and a person who answers the phone when a stranger’s model fails at three in the morning. Somebody has to own that. China assigned ownership to existing organisations — two central SOEs and one listed company — rather than chartering new ones. That is a cheaper institutional move than founding an entity, and it is also a more constrained one, for a reason that will matter below.

Second, and this is where the form becomes genuinely interesting: what the automotive base claims to have solved is not compute. It is the collective-action problem of sectoral infrastructure.

Neutral, does not build cars, capabilities fully open is a solution to a specific and well-documented market failure. Every automaker building its own AI stack pays the full fixed cost of a foundation layer and captures only its own share of the returns; the smaller the firm, the worse the arithmetic, so the technology diffuses at the speed of the largest balance sheets. A neutral party supplying compute, data and model as a service — with data ownership retained by the OEM and flowing only under a compliance framework — is a textbook club good: excludable, non-rival within the club, jointly produced. That is not a public-good claim and it does not pretend to be one. The line “中立不造车” is the credibility device, and it is credible only because the operator is structurally prohibited from competing downstream. Compare with the usual failure: the platform builder is also the largest user, so neutrality is a promise rather than an architecture. China Mobile, as a carrier, genuinely is not an automaker. The choice of host is the governance design.

The same logic exposes the constraint in “能力全开放”. Open to whom, on what terms, with what review? The published text says “the whole industry” and lists ten-plus state-owned automakers as adopters. It does not say whether a private automaker with a competing stack gets the same service level, whether a foreign-invested joint venture does, or whether a university lab gets access at all. This is not cynicism; it is the ordinary incompleteness of a first-year progress report. But it is precisely the distinction drawn yesterday between a cooperation output and a contribution artefact: an output is reported in participants and installations, and this document reports participants and installations. The aperture is stated for the compute — 112.7 EFLOPS, managed — and not stated for the access.

Third, the household-appliance base publishes the most institutionally sophisticated mechanism of the three, and it deserves a slow read.

“需求从行业中来、成果到行业中去”, executed through a supply-demand platform, open-call contracts (揭榜挂帅) for common problems, and an application service market for validated solutions. Strip the vocabulary and what remains is a procurement architecture for problem definition. The scarce input in industrial AI is not a model; it is an agreed statement of the problem that several firms recognise as theirs and will jointly pay to have solved. Firms do not publish their bottlenecks — a bottleneck published is a competitive disadvantage revealed. So the base interposes a broker: the platform collects requirements privately, distils them into a common statement, issues an open call, and returns a validated solution to every contributor. That is a real institutional design and it has a real open-source analogue: it is what a foundation does when it runs a specification process among competitors who cannot coordinate directly. The analogy is structural, not flattering. Open-call contracting is how you buy a commons problem when you cannot donate one.

But notice what the mechanism produces and what it does not. It produces a solution, validated and resold. It does not produce a repository, a licence, a maintainer or a contribution guide. The output of 揭榜挂帅 is a deliverable; the output of a foundation’s specification process is a document anyone may implement and nobody owns. Both are legitimate; they are not the same institutional object — and a system building one while calling it the other will keep reporting participants.

Fourth, the culture-and-tourism base published enough to be checked, and the check produces the day’s sharpest finding.

It named its datasets and their sizes: 40,000-plus ancient buildings, 400,000-plus components, 100-plus cultural spaces across 22 provinces. It named its models. It named its tenants. It did not state a licence, a deposit, or an access rule for a single one of those datasets. Everything in the announcement is a publication; nothing in it is a commons. A dataset a government aggregates can be opened cheaply once the aggregation exists — that is the whole advantage. But the three-year targets are 15PB of data, 33 scenarios, 15 standards: all counts of artefacts held, none of them a count of external users, outside contributors, or downstream repositories built on released data. The measure of a test environment is not its inventory. It is whether someone outside the base could run the test.

And here the difference between a Chinese pilot-test base and, say, a Hugging Face evaluation harness or a Linux Foundation conformance programme becomes precise rather than rhetorical. Those publish aperture — which datasets, which versions, which scoring, which exclusions — because their value to an outside developer depends entirely on the outside developer reproducing the result. A base whose value is to its own tenants has no such incentive: its tenants are inside the wall, and reproducing the test is an internal matter. Yesterday’s file made this point about download figures and it runs in both directions. Today’s version: a physically real, nationally chartered, richly furnished test environment can be, for every institutional purpose that matters to a stranger, opaque — and that opacity would not be a failure of the base. It would be a correct implementation of a different objective function.

Fifth, the reading has to cut against the easy version, because the easy version is wrong three ways.

First: the missing middle is real, and this is a serious attempt at it. The gap between a downloadable model and a working deployment is not a slogan. The FDE reporting from Zhejiang this week describes exactly that gap — a moulding plant that spent RMB 2 million-plus on an AI scheduling system which failed completely because one part carried three codes in the ERP and the capacity data was three years stale. No licence, no repository and no download count fixes that. A pilot-test base with a commissioning rig, a trusted data space and someone who knows the plant’s ERP does. The medical base’s reported outcome — a targeted-gene test from tens of thousands of yuan to a few hundred, in a minute — is the kind of number that only appears when someone runs the thing inside a hospital.

Second: openness is not absent; it is instrumental and specific. The healthcare opinions say “open” compute and “promote the building of medical AI open-source software.” The automotive base retains OEM data ownership instead of appropriating it. The appliance base’s neutrality is structural. These are genuine commitments to reduce exclusion, and reading them as theatre would be doing exactly what this series criticises when others do it in reverse. The commitments are real; what they commit to is a bounded openness — and saying so is a description, not an accusation.

Third, and least comfortable: the base form is not a Chinese peculiarity. NIST’s AI consortia, the EU’s AI Factories and testing-and-experimentation facilities, the UK AI Security Institute’s evaluations, Fraunhofer’s institute model, German and Japanese industrial testbeds — every industrial state builds entities to close the gap between paper and product, and most are neither open source nor open access. The difference is not that China built a base. It is what the base is asked to produce: a validated solution for members, or a reproducible test for anyone. That is an objective-function choice, and it is choosable.

Sixth, the frame that should be resisted. The temptation is to file this as “the state building infrastructure for open source” and stop. Read the primary documents and the relation inverts. The base is not an open-source instrument with a state wrapper; open source is one input among several to an industrial-conversion instrument. In the automobile, appliance and medical bases the operative words are pilot-testing, validation, standards, scaled promotion — the vocabulary of industrialisation, not of contribution. The culture base does not open a repository; it opens a campus. The appliance base signs ecosystem partners, not contributors. This is a coherent industrial policy, and calling it open source because some of its inputs are open models describes the inputs and mistakes them for the purpose. The base’s function is to convert downstream — from model to production line — which is precisely the stage at which an open licence stops doing any work.

And then the Wuzhen frame arrives tomorrow, and it does the framing for the state.

The 2026 WIC Wuzhen Summit runs 2026-11-02/05 under the theme “Open Source, Openness, Co-construction, Shared Benefits — Building a Community with a Shared Future in Cyberspace.” On 2026-11-01, one day before, the country’s first open-source-themed experience hall opens in the Wuzhen World Internet Science and Technology Museum — three floors, running to 2026-11-05, showing on the ground floor an “AI open-source major scientific and technological achievement exhibition” and an exhibition of the national AI application pilot-test bases; on the third, an “open-source parlour” hosting the Wuzhen Talk · Open-Source Youth Forum and Innovation Night, under a “365 days, never closes” operating mechanism.

The two halves of the same fortnight are one sentence. A national open-source experience hall opens, and the exhibits inside it are pilot-test bases. The public value is described by one Zhejiang official as making open source “no longer professional code for the geek circle, but a public resource everyone can see, touch and participate in.” That is the most honest sentence in the week’s coverage — it is a theory of what open source is for, and it is a theory of public legibility, not of public production. Visible, touchable, participatory is not the same as forkable, modifiable and redistributable. They are not the four freedoms, and they were never claimed to be.

The checkable question — three observables, each answerable from a public record, and the first two fall due within a month.

First: does any of these bases publish, before its first anniversary, the access terms for an outsider — not “the whole industry,” but a named class of user, a stated service level, an application path and a rejection path? “Capabilities fully open” becomes checkable the moment a private automaker, or a university lab competing with an incumbent, can be shown to have received compute on stated terms. One published access rule converts a cooperation output into an institution.

Second: does the culture-and-tourism base release one of its four named datasets under a licence a stranger can read? Its inventory — ancient-building genes, 40,000 structures, 400,000 components — is an aggregation no other party in the world holds. The plan counts 15 standards and 33 scenarios. A licence is neither, and its presence or absence is the cleanest single observable of whether the base intends a commons or a tenancy.

Third, and the one that generalises beyond China: when a state builds the test environment, what does it optimise — the number of validated solutions inside the wall, or the number of reproductions outside it? Both are defensible. Only one produces an ecosystem that survives its funder. A pilot-test base can industrialise a technology nobody else can replicate; a commons can be replicated by someone the funder has never met. The choice is not about money or capacity. It is about who is allowed to check the result.

Sources: