⚠️ 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 — 2026-11-08

🏗️ Institutional Change — From a Certificate of Contribution to a Condition of Employment: How Three Instruments Made Open-Source Work Measurable, and What They Agreed to Measure

1. The three artefacts, and why they belong in one entry. All three were published or executed in the same nine-day window at the end of the 2026 Wuzhen summit, all three are about the measurement of open-source contribution, and none of them is a licensing or governance document. Read together they describe a measurement infrastructure — and the infrastructure is converging on one unit of account.

  • Zhejiang University’s Taiyi 2.0 (太乙开源生态服务平台2.0), released 2026-11-04 at the “Internet Light” Expo open-source session of the World Internet Conference, hosted by the Zhejiang Provincial Department of Economy and Information Technology, with Zhejiang University’s School of Software Technology, its Ningbo International Science and Technology Innovation Center and four other units as co-organisers. The school’s own account published the release on 2026-11-06.
  • A national group standard for the OPC talent-capability model, project-launched at the same expo by Hangzhou Dianzi University (杭州电子科技大学), jointly with the Hangzhou Municipal Talent Group, the Xihu District Party Committee Organisation Department, the Zhejiang Association for Intelligent Economy and Smart Cities, Datawhale and SenseTime — reported by Hangzhou Net on 2026-11-06. Its draft version was published on 2026-06-08, and Xihu District had already opened a dedicated channel in its district-level “F-class” talent certification on the strength of it.
  • OpenAtom’s own talent-evaluation line, running since its Tsinghua session of 2025-04-18, where the foundation released “Campus Trip 2.0” (校源行2.0) built around “a contribution-oriented open-source talent evaluation mechanism” (以贡献为导向的开源人才评价机制), with 25 universities and research institutes forming the first pilot working group.

2. The policy that authorises all three, stated by the ministry in its own words. At the Tsinghua launch, Vice-Minister of Industry and Information Technology Xiong Jijun gave three suggestions, and the second is the operative one: “let open source be used in talent evaluation” — 让开源在人才评价中"用起来" — building “a contribution-oriented open-source talent evaluation mechanism”. The word is 用起来: use it. Not study it, not pilot it — put it to work inside the evaluation machinery. That single verb is why every artefact below is an instrument with a number attached rather than a statement of principle.

3. What Taiyi 2.0 actually does — five elements, four steps, and a data pipeline. The platform’s own description of its design is unusually explicit about the mechanism, so it is worth setting out in its own vocabulary. It is organised around five elements — 项目、任务、贡献、能力、人才 (project, task, contribution, capability, talent) — through a four-step path “problem — solution — recognition — output” (问题—求解—认可—输出). The four components:

  • 太乙悬榜 (Taiyi Open Board). Enterprises, open-source communities and research institutes post technical problems, R&D requirements and innovation tasks. Universities may convert suitable tasks into course cases, practical-training projects or graduation theses, and may organise faculty and students to attack common technical problems collaboratively.
  • Open-source competition (开源竞技). Students enter real projects through competitions or by taking board tasks, and are assessed on requirement analysis, option comparison, collaborative development and validation.
  • Contribution measurement (贡献评价). The platform “aggregates multi-dimensional data — code submissions, issue discussions, merge requests, community collaboration — and focuses on analysing the process and results of human–AI collaboration”, combining code analysis, software-engineering metrics, AI analysis and “scientific evaluation methods” to quantify project value, collaboration process, contribution quality and capability performance, producing a “verifiable, continuously updated capability profile” (可验证、持续更新的能力画像). Those records are offered as a basis for course-credit recognition, scholarship and award selection, and degree-output recognition (学分认定、评奖评优、学位成果认定).
  • 太乙真聘 (Taiyi Real Recruitment). Employers match on the basis of open-source contribution, competition record and board-task output; students turn projects, solved problems and results into what the account calls “a real résumé written in real contributions” — the “Taiyi open-source CV”.

4. The Zhejiang University–Huawei CANN agreement, and the exact sentence that defines the division of labour. The signing was executed by Ni Chao (倪超), head of the Taiyi open-source community and associate professor at the School of Software Technology, and Chen Min (陈敏), chair of the PMC of Huawei’s CANN open-source community. The stated model is verbatim: “universities organise the teaching, enterprises open the engineering scenarios, open-source communities participate in guidance and evaluation” (高校组织教学、企业开放工程场景、开源社区参与指导和评价), with operator development and model adaptation (算子开发、模型适配) given as the example tasks.

5. The OPC standard, and the nine levels — because the levels are the argument. The Xihu District standard is three tiers and nine levels: learner, builder, founder (学习者、构建者、创始人), subdivided into cogniser, tool user, scenario identifier, workflow builder, application builder, value deliverer, business-loop closer, system collaborator, ecosystem contributor. Every level is assessed differently, and the assessment methods are stated:

  • L1–L2 by quiz and by public data from market platforms;
  • L3–L5 by expert review, co-construction agreement, employer evaluation, or the ModelScope (魔搭) platform;
  • L6 by having received a real commercial order, banded at ¥1,000 / ¥10,000 / ¥100,000;
  • L7 by annual revenue from a completed customer-acquire–deliver–collect–repeat loop, banded from ¥100,000 to ¥10 million;
  • L8 by annual token consumption — hundreds of millions / tens of billions / hundreds of billions — as the measure of running a multi-agent remote team;
  • L9 by how many people you mobilised: events of 100 / 1,000 / 10,000 people, employment of 10 / 100 / 1,000 people, and co-created revenue of ¥1 million / ¥10 million / ¥100 million.

The district’s F-class channel takes code citation counts, community influence and open-source project contribution as its evaluation basis, and a successful certification carries a rent subsidy of ¥2,500 a month for five years, up to ¥150,000 in total. The district organisation department’s stated next step is to move the standard “up to a group standard and a provincial standard” — which the November launch at Wuzhen is.

6. What the universities say the mechanism has already produced. The figures below are institutional self-reports and are recorded as reported, not verified:

  • North University of China (中北大学), one of the first pilot universities, issued the country’s first university-level open-source contribution incentive policy in June 2025 — a points system that feeds scholarship ranking, postgraduate-recommendation bonus points (保研), graduate research-project applications, faculty title review and performance pay in one loop, with a ¥1,000–10,000 contribution scholarship, a ¥8,000/¥4,000 project grant for natural-science key/general projects, and 0.4 bonus credits for a first prize in an OpenAtom competition. It reports 1,200-plus people assessed and 89 projects on OpenAtom’s evaluation system, and a three-tier chain of “professor-adviser / graduate-supervisor / undergraduate-contributor”.
  • Fudan University, also a first-batch pilot, records that open-source contribution entered scholarship evaluation and postgraduate-recommendation composite assessment, with a separate expert-defence track for students whose contribution is recognised in domestic or international communities, and a first “Fudan Open-Source Pioneer” round in October 2025: 16 individuals and 10 teams applied, 5 students and 7 teams awarded.
  • East China Normal University reports that OpenRank, its contribution-influence algorithm built on the CESI Open Source Governance standard series, was listed by MIIT as a 2024 group-standard application exemplar (工信部科函〔2025〕19号), and that it is cited in CAICT’s open-source ecosystem white paper and CCAST’s global open-source report; the school has participated in OSPP for six consecutive years.
  • Huazhong University of Science and Technology states it has already put open-source contribution into postgraduate-recommendation and faculty evaluation.

Institutional Analysis & Commentary:

A measurement regime has been built, and it is converging on a single denominator — and it is not code.

Look at what each instrument actually counts. North University of China counts points. Fudan counts expert-defence recognition. Taiyi counts code submissions, issues, merge requests, community collaboration — plus the process and results of human–AI collaboration — and outputs a profile. The OPC standard counts orders, revenue, token consumption and headcount mobilised. These are not five variants of one metric; they are five different answers to the question “what is a contribution worth?”, and only two of them are about the artefact.

The OPC ladder makes the shift explicit and, to its credit, honest: from L6 upward, the unit of account is money. A senior AI practitioner, in this instrument, is one who has taken a real order; an elite one has run a revenue loop; an ecosystem leader has mobilised a thousand people for wages. Token consumption at L8 is the purest expression of it — a bill replaces a record. Whatever one thinks of that, it has a determinate institutional consequence, discussed below.

The other four instruments do the opposite: they monetise the credential, not the work. At Fudan and North University of China, the contribution’s real payoff is bonus points in a competition for scarce slots — postgraduate recommendation, scholarships, faculty titles. That is genuine, and it is what the mechanism was designed to do: give an outsider’s work standing inside a closed allocation system. But a metric that allocates scarce slots is a metric that will be optimised. North University of China’s own figure of 0.4 credits for a first prize is exactly the kind of number that generates a coaching industry, and the institution already specifies the competition list. The design question that follows is not whether gaming occurs — it will — but whether the instrument can distinguish a prize-winning submission from a maintained dependency. On the published descriptions, only Taiyi’s aggregation of merge-request and issue data across time even has the raw material to try.

Now the sentence in Taiyi 2.0 that deserves more weight than the rest of it: “focuses on analysing the process and results of human–AI collaboration.” In a framework where a student’s contribution is produced with an agent, what is being measured is not the patch but the direction of the student. That is a defensible thing to want to measure — it is the difference between knowing what a tool can do and knowing what to ask of it. It is also the hardest measurement problem in the entire document set, and the account gives no method for it beyond “combining code analysis, software engineering metrics, AI analysis and scientific evaluation methods”. A capability profile derived from data a university does not independently control, evaluated by a rubric it has not published, for a purpose (degree recognition) that carries legal weight — that is a new institution with an unaudited instrument, and the audit will be requested by the first student who is ranked by it and disagrees.

Three conclusions, offered as readings.

(i) The CANN agreement is the most institutionally interesting item in the week, because it names who decides. “Enterprises open the engineering scenarios, communities participate in guidance and evaluation.” Huawei’s CANN is an accelerator software stack whose external-contributor majority this series documented on 2026-10-04 from Techzine’s reporting (61 percent external contributors, 5,200-plus active developers a month). Embedding its operator development and model adaptation tasks into university curricula, with the community evaluating the output, converts a commercial ecosystem’s talent pipeline into an academic credential. That is efficient, it is honest about its mechanism, and it has a structural property worth naming: the enterprise determines which tasks count. A curriculum built on a vendor’s task board inherits the vendor’s roadmap. The university gains real projects and loses the ability to decide what a real project is — an exchange most engineering faculties would make, and one that should be recorded as an exchange rather than described as a collaboration.

(ii) The same structural move recurs: the state does not fund open source, it certifies contributors. Follow the money across all three instruments. Xihu District pays ¥150,000 per certified individual in rent subsidy. North University of China pays ¥1,000–10,000 per ranked contributor and 0.4 credits per competition win. Zhejiang pays for a platform that produces profiles. Not one of these payments is to a project, and not one is contingent on a repository being maintained. This series has argued since 2026-07-27 that China’s binding constraint on open source is that distributed contributions go unrecognised; the pilot programme is the answer to that objection, and the answer has taken the form of certifying people rather than of sustaining commons. That is a coherent choice — people are what a shortage-of-talent diagnosis predicts you will subsidise — and it means the question that will decide whether any of this produces durable software is the one no document addresses: who is paid to review the next pull request in year three.

(iii) The OPC standard’s monetary ladder is a labour-market instrument wearing a skills framework’s clothes, and it will be read as a price list. Publishing nine levels where the top ones are denominated in revenue, token spend and headcount does not merely describe a market; it tells every institution that adopts the standard what to reward. It is the same regulatory technique this series examined in MIIT Notice No. 42 on 2026-11-07: attach the binding force at the point where a licence or a subsidy is issued, not at the point where the choice is made. Here the licence is a district talent certification and the subsidy is rent. The falsifiable prediction is simple: within two years, the number of open-source practitioners certified under these schemes should greatly exceed the number of people whose commits appear in the projects they cite. If it does, these instruments measured a market for credentials; if it does not, they found a way to pay maintainers.

Checkable questions, all answerable from documents the institutions say they will publish.

(i) Will Taiyi’s capability profile be inspectable by the person it profiles? A student ranked by a system has, at minimum, the standing to see the inputs and contest them. Whether Taiyi exposes the underlying data and the rubric is the difference between an evaluation and a verdict.

(ii) Which open-source licences and which repositories does the OPC standard’s F-class channel actually verify? “Code citation count, community influence, project contribution” are all obtainable from platform metrics, and platform metrics are the cheapest thing in software to influence. Whether the district checks a contribution against a repository’s own history — commits, reviews, maintainer status — determines whether this measures work or measures visibility.

(iii) Does the OPC standard, once a group standard, get a published text? The June draft was reported through district and municipal media; a group standard is a document with a number and a drafting committee. The drafting committee’s composition is the answer to who wrote the ruler.

(iv) In the Zhejiang University–CANN curriculum, who grades — the course instructor, the community maintainer, or the platform? Three evaluators, three incentives. The agreement names all three as participants and assigns no final authority, which is the kind of gap that gets resolved by whoever shows up.

Sources: Zhejiang University School of Software Technology — 太乙 2.0 在乌镇发布!我院太乙开源社区与华为CANN开源社区签署战略合作协议 (2026-11-06) · Taiyi platform · Hangzhou Net — 杭电亮相互联网大会 牵头人工智能OPC人才能力模型团体标准研制 (2026-11-06) · Hangzhou Net / Beautiful Xihu — 如何识别AI开源人才?OPC九级能力标准(试行)正式发布 (2026-06-10) · Daily Business News via Hangzhou Net — 西湖区发布OPC九级能力标准 如何识别AI开源人才? (2026-06-11) · Chaoxi News via Tencent News — 如何识别AI开源人才?OPC九级能力标准正式发布 (2026-06-11) · OpenAtom Foundation — 开放原子"校源行"(清华站)成功举办,发布以贡献为导向的开源人才评价机制 (2025-04-18): Vice-Minister Xiong Jijun’s three suggestions; 25 pilot universities · OpenAtom Foundation — 中北大学以开源人才评价机制破局创新 (2025-07-01): points system, 保研 bonus 0.4 credits, ¥1,000–10,000 scholarship, 1,200-plus assessed, 89 projects · OpenAtom Foundation — 复旦新工科人才培养破局"唯绩点" (2025-10-16): 16 individuals and 10 teams applied, 5 students and 7 teams awarded · OpenAtom Foundation — 复旦大学计算与智能创新学院深耕开源人才评价改革 (2026-02-28) · OpenAtom Foundation — 华东师范大学构建一体化开源成长平台 (2026-02-04): OpenRank, MIIT group-standard exemplar 工信部科函〔2025〕19号 · OpenAtom Foundation — 校源行走进华中科技大学 (2026-04-28): contribution into postgraduate recommendation and faculty evaluation; Wuhan’s 2027 million-level talent target


🔬 Field Note — What a Vendor Forge Looks Like From Inside: 100 Commits of Ascend’s PyTorch Adapter

1. Why this artefact, and what it is not. Every Chinese project this series has examined at the contribution level — openEuler’s kernel on GitHub, on 2026-11-05 — was examined through a mirror. The PyTorch adapter for Huawei’s Ascend NPUs (Ascend/pytorch, a.k.a. TorchNPU) is a case where GitHub is not merely a mirror of the working repository but a repository that is actively committed to: pushed_at 2026-11-07T10:21:52Z, the day before this briefing. The sample below is the 100 most recent commits, from 2026-10-30 to 2026-11-07, pulled from GitHub’s public API. It is one repository, one week, and it is offered as a field note rather than as a finding about an ecosystem.

2. The author distribution, measured. Of the 100 commits, 51 carry an @huawei.com address, 16 carry an @h-partners.com address, 14 use @163.com, 12 use @qq.com, 3 use @126.com, 2 use isrc.iscas.ac.cn, and 1 each @hisilicon.com and @gmail.com. Two accounts are automation (pta-robot, ascend-robot), accounting for 15 commits, and they are the two largest single “authors” in the sample. That leaves 85 commits by 49 named individuals.

3. The sentence that has to be said carefully, because the naive reading is wrong. A qq.com address is not evidence of an external contributor, and saying otherwise would be exactly the kind of inference this section exists to refuse. Chinese enterprise engineers routinely commit with personal mailboxes, and h-partners.com is the domain associated with Huawei’s partner-labour ecosystem — the site of a large, contracted engineering workforce. So the honest aggregation is: at least 67 of 100 commits come from corporate or contracted mailboxes, at most 33 cannot be attributed to Huawei’s extended engineering organisation on the evidence of the commit header alone, and the true number of independent outside contributors is not determinable from this sample. What is determinable: 51 distinct author names appear, commit messages are mixed Chinese and English (fix(inductor): 修复 Fast GELU approximate 参数兼容性 next to feat(inductor): port Triton cumsum support to 2.13.0), and the engineering work is ordinary — Inductor backend ports, GIL-decoupling refactors, sanitizer dependency splits.

4. The number that matters more than the mix: the review surface. In the same window, Ascend/pytorch has 36 pull requests in total, ever, of which 12 were created since 2026-06-01 and zero since 2026-10-01 — against 11 issues created since 2026-10-01 and 123 open issues. The contributors’ code is not arriving as pull requests. It is arriving as commits on the default branch.

Institutional Analysis & Commentary:

A forge with commits but no review queue is not a commons; it is a release train with a public log. That is the whole field note, and the mechanism behind it is not mysterious. TorchNPU is a hardware adapter: it must track PyTorch’s upstream API, Ascend’s driver and CANN toolkit releases, and it ships to customers who buy the silicon. Its change cadence is set by a release calendar, not by a contributor’s interest. Under those constraints, review capacity becomes the scarce resource — and an organisation that has review capacity for its own engineers does not automatically acquire the ability to review a stranger’s patch. The 36-PR total is the measurement of exactly that: not hostility to outside contribution, but the absence of a queue with a service level at which outside contribution could be received.

The institutional read is that the binding constraint on this project is the same one identified in the Taiyi entry, viewed from the other end. The state’s instruments are building a supply of certified contributors — 1,200 assessed at one university, 25 pilot institutions, a nine-level ladder, ¥150,000 per certified individual. This repository is what the demand side looks like: a project that publishes code promptly, reviews outside patches rarely, and absorbs work through an internal engineering organisation. Certifying contributors faster than the projects they would contribute to can review them produces a credential surplus, not a codebase. The remedy is unglamorous and entirely internal to the vendor: a public review queue, a stated response time, and a maintainer list that includes people who do not work for the company.

And the counter-observation, which belongs in the same paragraph. A project that commits openly on a public forge is doing something the closed alternative does not: the history is inspectable, forkable and attributable, including the automated commits and the contracted-labour addresses. That is a real commitment, it is the precondition for everything else, and it is the reason this analysis is possible at all. Openness of the log and openness of the decision are two different goods, and China’s hardware-software stacks have largely secured the first while the second remains internal. The useful next measurement is not contributor counts; it is interaction lag on outside issues — how long between an issue opened by a non-employee and a maintainer response. That number is available, and its trend would say more than any proportion of email domains.

Checkable questions.

(i) Is GitHub the working repository or a mirror? The README and contribution guide say; the answer determines whether the 36-PR total is a review-capacity fact or a mirror artefact. If it is a mirror, the question moves to the upstream forge, whose review behaviour should be measured the same way.

(ii) What is the median time-to-first-maintainer-response on issues from non-Huawei addresses? This sample contains issues from clearly external accounts (weifulun, 202AI-cell, VenusTZZ, kleinblueworld) filed between 2026-10-20 and 2026-11-02. The lag distribution across those is directly computable and would settle the substantive question.

(iii) Does any published maintainer list include a non-employee? As with the company-donation analysis this series has run on Alibaba’s projects, the appearance of a non-originating company’s name in CODEOWNERS or a maintainer file is the single clearest institutional indicator that a project has crossed from vendor code into a governed commons. For a hardware adapter the bar is higher — but the indicator still works.

Sources: GitHub API — Ascend/pytorch repository metadata (604 stars, 106 forks, 123 open issues, pushed 2026-11-07) · GitHub API — 100 most recent commits, 2026-10-30 → 2026-11-07 · GitHub — Ascend/pytorch pull requests (36 total; 12 since 2026-06-01; none since 2026-10-01) · GitHub — Ascend/pytorch issues (open issues filed 2026-10-20 → 2026-11-02) · This series’ 2026-10-04 briefing — Techzine Global reporting CANN’s external-contributor majority (61 percent) and 5,200-plus active monthly developers


🔍 Commentary — The Credential Is Now the Instrument

First: this week’s two entries are one mechanism read from two ends. A university system that converts contribution into credits, ranks and rent subsidy; a vendor forge that receives contribution through an internal engineering organisation. The first is supply, the second is demand, and the two are being built at visibly different speeds. The supply side has a ministry, a foundation, 25 pilot universities, a municipal talent office, a nine-level standard and a subsidy schedule denominated in yuan per person. The demand side has a review queue that outside code reaches roughly never. No amount of certification fixes that mismatch, because certification is not the scarce factor. This is the recurring error of an ecosystem built by authorisation rather than by coordination — and it is the error this series keeps landing on: the instrument that is easy to build gets built, and the instrument that requires someone to answer a stranger’s email does not.

Second: the interesting thing about the Taiyi/OPC designs is that they are honest about what they measure, and the honesty is the problem. The OPC ladder says out loud that a senior practitioner is someone who has taken an order, and that an ecosystem leader has moved a thousand people into wage employment. There is no hypocrisy to attack there; that is a labour-market framework with the labour market left in. But an open-source commons is precisely the institution that produces value nobody can price at the moment of production — a maintained dependency, a documented bug, a review of someone else’s patch, a fork that kept a project alive. A measurement regime whose top levels are denominated in revenue and token spend will, over time, select for people who sell rather than people who maintain. The remedy is not to remove the money bands; it is to make sure the lower bands can see a maintainer’s work as clearly as they see a vendor’s order. On the documents available, that is the missing row in the table.

Third, a caution that runs against this series’ own enthusiasm. It would be easy, and cheap, to read all three instruments as “measurement theatre” — credentials manufactured to fill a quota. The evidence does not support that. North University of China’s 0.4 credits moved a real selection mechanism; Fudan built an expert-defence track and ran an award round with 16 individual and 10 team applications, which is a small, expensive, non-theatrical process; Taiyi is a working platform at a public URL with a data pipeline described in operational detail; and the OPC standard was used by a district government to allocate real money before it was ever a group standard. These are institutions, not press releases. The criticism that survives contact with the documents is narrower and more useful: they certify practitioners, they do not fund maintenance, and they have not yet published the rubrics by which either is judged.

Fourth, the sentence worth keeping — and it comes from the ministry. “Let open source be used in talent evaluation.” Two years ago the objection to Chinese open source was that contribution did not count. It now counts: in credits, in bonus points, in rent subsidy, in a nine-level ladder and in a capability profile. The next question is the one that a counting system always raises and never answers by itself — whether what now counts is what matters, or whether it is what can be counted. The only way to know is to look at the repositories in five years and count who reviewed the pull requests.