⚠️ 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 — 2026-10-20

🏛️ Policy & Regulation — GB/Z 185—2026 and Its Reference Implementation: A Seven-Part National Standard, a Community Built to Carry It, and Zero Pull Requests After Three Months

1. The standard, and what kind of instrument it is. On 2026-06-26 the State Administration for Market Regulation (SAMR, 市场监管总局) announced the publication of the 《人工智能 智能体互联》series — seven national standardisation guidance documents (国家标准化指导性技术文件), registered as GB/Z 185—2026. The seven parts cover, in the announced design logic: Part 1 overall architecture; Parts 2 and 3 agent identity code and identity management (the “digital identity card”); Part 4 agent description (capability profile); Part 5 agent discovery (supply-demand matching); Parts 6 and 7 agent interaction and tool invocation. Zhu Meina (朱美娜), deputy director of SAMR’s standards technology department, stated at the press conference that the series was issued as guidance documents rather than mandatory standards precisely because this is the “agile standardisation arrangement” for the industry-cultivation phase — maximising compatibility with multiple technical routes, reserving room for trial and error, with the explicit plan that identity-code standards will later be converted to mandatory national standards, alongside new work on agent audit and agent transactions.

2. The implementation chain the issuing bodies described. Fan Kefeng (范科峰), vice-president of the China Electronics Standardization Institute (CESI, 中国电子技术标准化研究院), set out a “standard co-research, open source, adaptation assessment” (标准共研、开源开放、适配评测) three-in-one landing model with three named components: (i) a “Pioneer Programme” (先锋计划) launched through the national AI standardisation sub-committee TC28/SC42, with “one hundred-plus key ecosystem enterprises and a large number of individual developers” doing pilot verification; (ii) CESI “jointly with Beijing University of Posts and Telecommunications (BUPT) and other academic research forces” leading AIP (Agent Interconnection Protocol) open-source code development and iteration; (iii) an assessment toolkit built on the “Qiusuo” (求索) national AI standard assessment benchmark system. Separately, Fan Bin (范斌) of MIIT’s science and technology department reported that the series was drafted by 70-plus leading enterprises, universities and research institutes and over 100 experts, absorbing 600-plus public comments, that a “Joint Initiative on Co-construction and Sharing of Agent Protocols” (《智能体协议共建共享联合倡议》) had been launched, that 100-plus enterprises had joined the initiative, and that 50-plus enterprises were running pilot applications.

3. What “open source” produced, measured. The AIP project was publicly released on AtomGit on 2026-07-07 as v2.1.0, described by the OpenAtom Foundation as “the companion code tool synchronously formed by the seven national standardisation guidance documents,” and reported as having passed 6,000 downloads. On 2026-10-20 this briefing inspected both the AtomGit organisation aip-pub and the operating zone aip.openatom.tech (hosted by OpenAtom, built by CESI and BUPT). The measured state:

  • Organisation aip-pub: 9 repositories, 40 stars, 10 forks, 12.25K downloads, 3 issues, 0 pull requests.
  • Operating zone aip.openatom.tech: 156 members, 20 repositories, 6 collected repositories, 44 cumulative stars, and the platform’s own counter reads “累计 PR 数 0” (cumulative pull requests: 0), “累计 Fork 数 0”.
  • Repository status: 11 of the 12 listed organisation repositories are marked 已归档 (archived), the most recent archival date being 2026-06-22. The only non-archived repository is ACPs-community — “Agent Interconnection Protocol Project based on the GB/Z 185 standard” — with 10 stars, 8 forks, 1 issue, 3 project members, last updated 21 days ago (around 2026-09-29).
  • Community dynamics on the zone: the most recent open discussion is dated 2026-09-02 (“please help us design our platform users’ agent identity code”); the one before that is dated 2026-07-13.

Institutional Analysis & Commentary:

The headline number is not the zero. The zero is the finding, and it is worth stating precisely: after a standard with 70-plus drafting organisations, 100-plus experts, 600-plus absorbed comments, a 100-plus-member co-construction initiative and 50-plus pilot enterprises, the accompanying open-source implementation has recorded no pull request from anyone outside its three project members across its entire public life, and eleven of its twelve repositories have been archived.

That fact requires an explanation that does not assume bad faith, and the honest one is structural. A reference implementation of a standard is not a commons. It is a conformance fixture. Its purpose is to demonstrate that the clauses in GB/Z 185—2026 can be instantiated in running code, so that enterprises writing their own agents have an executable definition of conformance. A conformance fixture is correct when it matches the specification, and complete when the specification is stable. Both properties are hostile to outside contribution: any external patch risks divergence from the text, and the text is the authority, not the repository. In this respect AIP is doing exactly what it was built to do — it is a test oracle with a git history.

But there is a second, harder observation. When the issuing bodies describe their landing model as “standard co-research, open source, adaptation assessment,” the word “open source” is being used to name a phase in a standards process, not a governance form. The three parts are sequential and administered: research the standard, publish code that exemplifies it, assess products against it. In that sequence no party holds a right that a codebase would grant — no right to fork and have the fork be legitimate, no right to have a patch reviewed on technical grounds by someone who does not answer to the standard’s drafting committee, no right to elect the reviewers. Those rights are what distinguishes the production commons from the reference implementation, and their absence is legible in the archive timestamps: the repositories were archived on 2026-06-22, four days before the standard was announced on 2026-06-26. The code was finished when the text was finished, because the code’s job was to prove the text.

Compare the structural alternative. When the W3C produced HTML5, the specification and the browser implementations were separately owned by parties with divergent interests, and the test suites existed so that independent implementers could check themselves. When IETF working groups produce a protocol, the “reference implementation” clause exists in some charters and is controversial in most, precisely because participants understand that a reference implementation written by the spec’s authors confers an advantage on those authors. GB/Z 185—2026 has gone further than most: it has a reference implementation, an operating zone, a SIG structure, a co-construction initiative, and an assessment regime, all authored within one institutional family — CESI writes the standard, CESI co-leads the code with BUPT, CESI’s “Qiusuo” platform does the assessment, and OpenAtom hosts the zone while its deputy secretary-general sits in the same forum announcing project hosting and IP protection as the foundation’s contribution. That is a complete vertical chain, and it is defensible on efficiency grounds. It is simply not the institutional form that produces the thing open source historically produces, which is a codebase whose direction can be changed by people who were not in the room when it was specified.

One checkable prediction. Zhu Meina’s stated plan is that identity-code standards convert to mandatory national standards, and that agent audit and agent transaction standards follow. If that conversion happens while the accompanying code remains an archived fixture, then Chinese agent interconnection will have achieved interoperability by legal compulsion rather than by adopted protocol — and the difference will be observable in a very specific place: whether non-Chinese agent frameworks ever implement AIP, or whether “interconnection” ends up meaning interconnection among the parties who were in the drafting room. The 12.25K downloads against 0 pull requests already suggests which of those two futures the code is optimised for. Downloads mean people are reading the fixture. Zero pull requests mean nobody believes they can change it.

Sources:

⚖️ Legal & Licensing — Six Reports, One Sub-Forum, One Incubated Project: What CESI and Mulan Put on Stage in Beijing on 2026-10-13, and What the Xingzhu/Clink Launch Says About Who Owns an Open Interconnect Standard

1. The forum and its sponsor. On the afternoon of 2026-10-13, at the Beijing Conference Centre, the “Open Source Standards and Ecosystem Construction” sub-forum (开源标准与生态建设分论坛) was held as part of the 2026 New-Industry Standardisation Navigation Forum (2026新产业标准化领航论坛). It was convened to address “open-source standards-system construction and local open-source ecosystem cultivation,” and gathered “competent authorities, research institutes, the OpenAtom Foundation, leading enterprises and open-source community representatives.” Speakers: Li Yan (李琰), deputy director of MIIT’s Information Technology Development Department; Fan Kefeng (范科峰), vice-president of CESI; Li Bo (李博), deputy secretary-general of OpenAtom. Chairs: Su Wei (苏伟), deputy director of CESI’s Software Centre, and Yang Liyun (杨丽蕴), head of CESI’s Open Source Research Office. Note the identity of the host organisation: the same CESI that authored GB/Z 185—2026 and co-leads the AIP code is also the body convening the forum on open-source standardisation, and its own staff chaired the sessions.

2. What Fan Kefeng and Li Bo actually committed to. Fan stated that “China’s open source has entered a new stage of scale development,” and that CESI, under MIIT guidance, “systematically advances open-source standardisation and open-source ecosystem construction.” Li Bo stated that the foundation will “continue to play its platform-carrier role, doing well in open-source project hosting, intellectual property protection, and community governance capacity building,” will strengthen linkage with “competent authorities, standardisation bodies and industry parties,” and will “open up the conversion path from standard text to project operation (打通标准从文本到项目运营的转化路径)”, while “gathering enterprises, universities and developers to cultivate local open-source culture.”

3. The incubated project — Xingzhu, and the Clink protocol behind it. The forum held a launch ceremony for Mulan Open Source Community’s newly incubated Xingzhu project (星杼项目). Per the official account of Mulan: the project is “led and co-built by ZTE (中兴通讯主导共建)”, aims to “accelerate the landing of the Clink interconnect protocol standard (加速Clink互联协议标准落地)”, and is intended to “lower the development and adaptation barrier for multi-vendor GPU interconnection (降低多厂商GPU互联开发适配门槛)”. It will “rely on Mulan Open Source Community’s governance and operations capability, gradually open technical capability to industry, and inject new momentum into the computing-power open-source field.” ZTE’s open-source director Meng Wei (孟伟) was listed among the speakers.

4. What is publicly knowable about Clink’s institutional status. Clink is an interconnect protocol under MIIT guidance with ZTE as lead drafter. As of 2026-07-03 the public state, per a market-discussion thread that this briefing treats as unverified investor commentary and not as a source of institutional fact, was that Clink was “still at the standard-drafting stage,” would not be a mandatory standard, and that “even if the final version matures it will take at least several years.” The same thread described it as a scale-up interconnect intended for use by domestic chip vendors’ packages, framed by that commenter as “effectively excluding foreign links.”

5. The other five reports, and the one that matters most here. Yang Liyun presented SJ/T 12304-2026, the electronics-industry standard on grading the openness of large-model technology systems (大模型技术体系开源开放分级标准), addressing “blurred openness boundaries and frequent licence disputes,” and introduced the L1–L5 completeness and O1–O4 openness assessment framework plus Mulan’s “AI-OpenG” assessment practice, positioned as “a standardised tool for evaluating and compliance-operating large-model open-source projects.” Wang Heshu (王荷舒), head of OpenAtom’s legal and IP department, presented on judicial developments and rule analysis for large models, drawing on the September 2026 《关于依法审理涉人工智能纠纷案件的意见》 (Opinions on the lawful adjudication of AI-related disputes) and domestic and foreign case law, and framed the courts as interacting with industry practice and standards to “jointly explore and construct new rules for the AI industry.” The remaining three: Wang Xu (王旭, Ant Group, Mulan TOC) on Agentic AI and the argument that “natural language is gradually becoming the new source code,” and on AI-agent participation causing PR backlogs and declining decision efficiency; Zhang Jinnan (章津楠, MetaX/Muxi) on AI-Infra open source and the value of interface and compatibility standards; Tang Bin (唐滨, Beijing Kaiyuan Industrial Software Research Institute, in preparation) on industrial-software open source; Li Yongle (李永乐, Huawei Computing, Mulan TOC) on openEuler in the agentic era, observing that as agents arrive “both contributors and users of open-source software change from people to people and AI”, that development cost falls while selection and maintenance cost may rise; Du Shaohui (杜韶辉, China Southern Power Grid Digital Grid Research Institute) on standards-led open-source construction in the electric-power sector, where standards govern project admission, code governance and security risk control.

Institutional Analysis & Commentary:

Three things in this forum are institutionally new, and they are all in the same direction.

First, the open-source standard has become an object of administration rather than an output of a community. Yang Liyun’s framing of SJ/T 12304-2026 is the clearest available statement of the Chinese model’s core problem and its proposed solution. The problem is stated honestly: “large models’ openness boundaries are blurred and licence disputes are frequent.” The solution is a grading scale — L1 to L5 for completeness, O1 to O4 for openness — with an assessment service (AI-OpenG) attached. This is a real contribution to a real mess; OSI’s Open Source AI Definition and the model-licence debates around Llama, Gemma, DeepSeek and Qwen have produced exactly the kind of boundary confusion 12304 is trying to grade. But observe what a grade is: it is an externally assigned rank, not an internally enforceable right. A model rated O2 rather than O4 has been described, not constrained. Nothing in a grading scheme lets a downstream user compel anything of the publisher, which a licence does. The distinction matters because grading is cheap to administer and cheap to comply with cosmetically, while licence obligations are expensive and create liability. A field that grades openness instead of enforcing it can drift, grade by grade, toward definitions that suit the graded.

Second, the forum’s most candid technology observation came from the people closest to the code, and it describes a governance problem they are treating as a research question. Wang Xu reported that AI-agent participation in collaboration is producing PR backlogs and declining decision efficiency. Li Yongle reported that contributors and users are becoming “people and AI,” that development cost falls while selection and maintenance cost rises. These are not marketing claims; they are the same observations that have dominated Western agent-tooling discussion. What is institutionally interesting is the response: both framed the consequence as new topics for standardisation research (“new standard research propositions,” per Li). The alternative response — that the bottleneck is review capacity, and that review capacity is allocated by maintainers whose authority derives from having done the work — would point toward strengthening committer institutions rather than toward writing standards about agents. A standard can specify how an agent submits a patch. It cannot manufacture the human who is willing to read it.

Third, and most structurally: the Xingzhu/Clink launch reveals the direction of the arrow between standards and open source at the hardware layer. Mulan’s announcement is precise about the causal order: the project exists to accelerate the landing of the Clink protocol standard. Not “the standard emerged from the community’s work,” not “the community converged on an interconnect,” but an already-drafted, MIIT-guided, ZTE-led standard is now to be housed in a community whose role is governance and operations capacity. The community is being used as a deployment vehicle for a standard, which is the exact inversion of the historical UNIX/Linux/POSIX sequence in which implementation ran ahead of specification and the spec later ratified what practitioners had already converged on. In the Mulan/Clink sequence, the specification exists first, its author is a single vendor, and the community arrives after to lower adoption friction for “multi-vendor GPU interconnection.”

This is not necessarily a criticism, and the honesty of the framing deserves acknowledgement. The problem Clink addresses is real and expensive: multi-vendor GPU interconnect in China is genuinely fragmented, every accelerator vendor ships its own scale-up fabric, and the adaptation cost falls on every downstream integrator. A shared interconnect at that layer would be a public good in the strict sense — non-excludable in use, and expensive to produce privately. The question is not whether the goal is good. It is who holds the rights that make a standard into an open one, and that question has a specific answer in a standard whose drafting is led by one company: the rights sit with the drafter, and the community’s accession adds adopters, not claimants.

Elinor Ostrom’s account of what makes a shared resource durable is not chiefly about who writes the rules; it is about whether the people bound by the rules participated in making them, and whether they can revise them without permission from above. Mulan Open Source Community has an established TOC and real operational discipline — that is why the project is being placed there, and the forum’s own speakers sit on that TOC (Wang Xu of Ant, Li Yongle of Huawei). The institutional question for Xingzhu over the next year is therefore answerable and narrow: does the project’s charter give non-ZTE participants the ability to change the protocol, or only the ability to implement it? If the TOC can amend Clink on technical grounds against the lead drafter’s preference, then the standard has been genuinely opened and Mulan has done something no other instrument in this series has managed. If the TOC’s role is operational — releases, CI, documentation, events — then Chinese GPU interconnect will have a competent community attached to a vendor-controlled specification, and the “multi-vendor barrier reduction” will have been achieved by making every vendor implement ZTE’s design rather than by letting them contest it.

A caution about the sources. Two of the six reports here — the L1–L5/O1–O4 framework and the AI-OpenG assessment — are described only through conference-side reporting; the SJ/T 12304-2026 clause text is not public. Clink’s drafting status rests on unverified investor-forum commentary from 2026-07-03, and the Xingzhu/Clink relationship is stated only in Mulan’s own launch announcement. What this briefing can establish is the sequence and the announced division of labour. What it cannot establish is the charter, and the charter is where the answer lives.

Sources:

🔍 Commentary — Two Ways to Institutionalise a Protocol, and the Reason the Difference Is Now Measurable

Two documents from the same fortnight, both about “open source standards,” both authored inside the same institutional family, and they differ on the one dimension that determines what happens to a protocol after its authors lose interest.

The first is GB/Z 185—2026, the agent-interconnection series. It has a real standard, a real reference implementation, a real operating zone, 156 zone members, 100-plus enterprises in a co-construction initiative, 50-plus pilot enterprises, and an assessment regime. It also has zero pull requests, eleven of twelve repositories archived, and a community whose most recent open discussion is a request for help designing an identity code — dated 2026-09-02, addressed to “各位老师” (respected teachers), which is the register of a seminar, not of a project.

The second is Xingzhu/Clink, launched into Mulan on 2026-10-13 to make a vendor-drafted interconnect protocol land across multiple GPU vendors. Its charter is not public. It may well turn out to be the more open of the two, because Mulan has an elected TOC that already contains people from Ant and Huawei — parties whose interests in an interconnect are not identical to ZTE’s.

The distinction I want to name is not “Chinese standards are bad.” It is older and applies everywhere, including to the Linux Foundation and to the W3C: there are two ways to make a protocol, and only one of them survives its author. The first is to publish a specification and a conformance fixture and invite the world to implement. The world will do so if the specification is good, and the resulting interoperability is real, and it will last exactly as long as the author keeps editing the fixture. The second is to hand over the right to change the thing — the amendment procedure, the technical veto, the reviewer appointment — to a body the author does not control. The first produces interoperability; the second produces portability, meaning the ability of the ecosystem to keep the protocol alive after its originator stops caring.

For agent interconnection, the question of which one we are watching has a brutally practical test, and it is not philosophical. When GB/Z 185’s identity-code standard converts to mandatory, what happens to a foreign agent framework that implements the protocol correctly but does not implement the Chinese identity-code authority? If the answer is “it interoperates,” then the fixture did its job and the standard is genuinely open. If the answer is “it does not,” then “open source” was describing the delivery mechanism, and the interoperability achieved is the kind you get inside one jurisdiction by administrative means. That test cannot be run today. It will be run, and the fact that it can be run at all is the useful thing — the 0-pull-request counter on a standard’s own community is an unusually honest instrument, because it measures the only thing that distinguishes a commons from a distribution channel: whether strangers believed they could change it.

Lin Clark’s observation about standards applies here without modification: the hard part was never inventing the format, it was getting everyone to agree. Agreement that is produced by graded assessment, pilot programmes and a mandatory conversion plan is not the same achievement as agreement that is reached by implementers who could have refused. Both are agreement. One of them requires the state to keep paying attention, and the institutions of the other keep working after everyone in the room has left.

Sources: