⚠️ 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 — 2026-10-24

🏛️ Policy & Regulation — The Registration Path Was Walked: What DeepSeek’s and MiniMax’s Completed Formalities Do to the Exemption

1. The report, and who issued it. On the evening of 2026-10-22, Yuyuantantian (玉渊谭天) — the new-media account operated by China Central Television (CCTV) — reported that DeepSeek and MiniMax, “two leading AI enterprises”, have both 顺利履行登记手续 (successfully discharged the registration formalities) and 完成开源和海外授权使用 (completed open sourcing and overseas licensing use). The report was carried the same night and the following morning by Yicai (第一财经), Sina Finance, and chinaz, in each case credited to Yuyuantantian rather than to the companies, and in each case with the same three nouns: registration discharged, open sourcing completed, overseas licensing completed.

2. The framing the state media chose, and the sentence that is doing the work. The report exists to answer a question the draft had raised. Its framing: “After the Catalogue appeared, some people raised questions — China emphasises technological openness, so why suddenly add ’export restrictions’ to AI? Is this a signal of a policy turn?” The answer given: the revision is not aimed at simply restricting outward export of AI technology but rather at “establishing clear boundaries” (划定边界) and at “normative management of exports of strategically significant technology.”

The operative sentence, which the trading press picked up verbatim and which this briefing treats as the finding, is this one:

商务部在相关答问中，对"什么不管"的解释，明显比对"什么要管"的解释更多、更具体。从这也能看出，《目录》的核心是划定边界，而非单纯限制AI技术对外出口。

(“In MOFCOM’s related Q&A, the explanation of ‘what is not managed’ is visibly longer and more specific than the explanation of ‘what is managed.’ This shows that the core of the Catalogue is to draw boundaries, not simply to restrict the outward export of AI technology.”)

The Yicai write-up of the same report adds two items of policy provenance worth recording: the revision had been in preparation for some months, with draft-consultation work beginning in the first half of 2026 involving research institutes, industry organisations and leading enterprises; and legal practitioners who participated report that “the control points for AI-sector technology entries have been continuously adjusted as the industry develops, and the comments and concerns that enterprises raised in those meetings are well reflected in this policy as published.”

3. The two preconditions the report itself states. These matter because they define the boundary of the claim. First: “判断AI产品’出海’是否属于技术出口管理范围，核心在于是否涉及限制类技术从中国境内向境外转移” — whether an AI product’s outbound movement falls inside technology-export management turns on whether a restricted technology is being transferred from inside China to outside it. Second, and directly quoted from the MOFCOM spokesperson’s 16 October Q&A: even when the underlying model behind an API, web application or SaaS product is covered, providing service to overseas users via API calls, web applications or SaaS is not subject to export restrictions. What the draft does reach is overseas deployment of a restricted model, with the facilitation channel being a Chinese company’s own overseas self-built, self-operated compute centre (境外自建自营算力中心).

4. The context the report supplies that a foreign reader will miss. The Chinese write-up positions the instrument explicitly against the US. Its account: the United States has, for several years, built export restrictions around advanced AI chips and related technology, “dividing control thresholds by country, region and specific entity”, in order to “maintain its own competitive advantage” by targeted bans and technological blockade; China’s approach, by contrast, is “to manage risk and promote international cooperation through clear rules.” It notes that this revision is an ordinary periodic update of the Catalogue, citing the 2020 addition of algorithm technology as precedent. And it identifies the regulatory focus as three risk classes — data security, sensitive uses, and compute base (数据安全、敏感用途、算力底座) — which “manages the underlying logic of the technology.”

Institutional Analysis & Commentary:

Yesterday’s briefing ended with a checkable question: what does the exemption actually cover, and does the implementing guidance define “open sourcing” narrowly or broadly? That question has not been answered by text. It has been answered — for now, and by demonstration — by the existence of two completed registration files.

The distinction that matters is between a gate and a toll booth, and this week the difference became concrete. A gate is an instrument that can say no. Its administrative cost sits in the refusal, and its information value is a by-product. A toll booth is an instrument that records every vehicle and refuses almost none. Its administrative cost is near zero per transaction, and the record it generates — who passed, carrying what, to whom, for how much — is the entire product. Announcement No. 48 was written with a gate’s vocabulary (prohibited, restricted, licensed) and a toll booth’s mechanism (register before export, automatic licensing, immediate release on filing). What the 2026-10-22 report establishes is not that the gate is open. It is that the traffic is flowing through the booth, has been counted, and the state has chosen to publicise the count. Two of the four or five Chinese labs whose models plausibly cross the 2-trillion-parameter or 10²⁶-FLOP thresholds walked the path within six days of the regime taking effect and emerged with a documented completion. In a system where the first mover’s experience sets the reference for everyone else, that publication is itself the policy.

Note what the state media chose to emphasise, because the emphasis is a design signal. “The explanation of what is not managed is longer and more specific than the explanation of what is managed.” That is an unusual sentence to put in a state outlet’s own voice. It is an argument about ratio of attention, offered as evidence about intent: a restriction instrument that spends most of its explanatory budget on exemptions is announcing that exemption is the normal case. Whether or not that inference is sound as textual analysis, it is being made deliberately and in public, which is what a implementing body does when it needs firms to keep publishing. A confidential licensing regime does not publish its exemption reasoning; a regime trying to preserve a release pipeline does.

The register, read against Ostrom rather than against export-control law, becomes more interesting than a border instrument. What has been built is not primarily a filter on goods but a stock of information about the flow of capability, collected before the fact, at the moment when the flowing party still has an incentive to be accurate. The four registration triggers documented earlier — ownership transfer, outbound licensing, change of controlling owner to a foreign entity, and anything MOFCOM later designates — are all moments when information about a technology’s disposition exists inside a Chinese legal person and would otherwise become unobservable once it crosses. Requiring the filing converts each of those moments into a record. That is the same informational move as a land registry or a ship-registry: not to prevent transfer, but to make transfer legible. And like every registry, the institutional question is not whether it records but who may inspect the record, and whether the recorded party can contest an entry — neither of which the draft specifies.

The honest counterweight, which this week’s reporting does contain and which should be given full weight. The report’s own premise is that AI firms were worried. Those worries were put to the ministries in consultation meetings and, according to the lawyers quoted, “well reflected in this policy as published.” That is a description of voice — affected parties being heard before the instrument was fixed — and voice is precisely the variable this series has repeatedly found missing from the administrative instruments it has examined. If the account is accurate, the drafting process for Announcement No. 48 included an actual consultative channel with the regulated firms, over a period of months, and the published draft reflects it. That is a materially different process from a plan issued with targets and no comment period, and it deserves to be recorded as such. The caveat is symmetrical: the voice quoted is that of firms and their lawyers, not of the downstream recipients in other jurisdictions who are the objects of the control, and not of any party who might wish to contest a decision to restrict.

The comparative point the Chinese report makes is not wrong, and this series should say so plainly rather than only noting its rhetorical function. The US framework it describes does have a classified threshold, a text that has not been published in full, and pre-release review access granted and withdrawn at executive discretion. Announcement No. 48 publishes its thresholds (2 trillion parameters, 10²⁶ FLOPs, top-ten placement on public evaluations judged at the time of export), opens a comment period with an email address and a deadline of 2026-11-15, and writes its exemptions into the operative text. A regime that states its thresholds and its exclusions, in public, with a comment period, is a better instrument on the dimension that matters most to anyone trying to plan — not because it restricts less, but because it restricts predictably. That is a real procedural achievement and it is independent of whether the substantive coverage is wise.

The checkable question. The report establishes an outcome; it does not establish a rule. Three observables follow, and all three are public. First: are the registration records or their aggregate published? The essence of a toll booth is the count. If MOFCOM or state media periodically report how many registrations were filed and how many resulted in releases, the regime is functioning as a monitored channel and firms will treat the filing as routine. If the count is never published, the same mechanism functions as a discretionary dossier collection, and the “success” of DeepSeek and MiniMax becomes an anecdote about a friendly window rather than evidence about a path. Second: does the next threshold-crossing release — a model at 2 trillion parameters, not at 428B — go through in the same six-day pattern? DeepSeek’s and MiniMax’s completed filings concern their existing releases and overseas licensing; the harder test is the frontier release made after the guidance is final. Third: what happens on 2026-11-15? The comment period closes. If the published final text retains the public-domain exclusion in the same words, the precedent is set. If the exclusion is narrowed in the final revision while the registration system — already in force — continues, then the clause this series read as an exit becomes a transitional passageway, and the firms that published inside the window will have made the right call for a reason they did not fully control.

Sources:

⚖️ Legal & Licensing — 428 Billion Parameters, Released the Morning After: What MiniMax’s M3.1 Flash Separates, and What the Market Priced

1. The release. On 2026-10-23, MiniMax (稀宇科技) announced the open-sourcing of the M3.1 Flash model weights — 428B total parameters, 23B active — publishing on Hugging Face (MiniMaxAI/MiniMax-M3.1-Flash) and ModelScope (MiniMax/MiniMax-M3.1-Flash). Two performance facts were published with it: the model handled more than 100 trillion tokens of real user traffic within 72 hours of going live, and cost fell 70 percent relative to the previous-generation M3, with capabilities “comprehensively improved.” A named ecosystem accompanied the release: vLLM, SGLang, BAAI’s FlagOS, Hugging Face and ModelScope are described as jointly driving the model’s adoption across different chips, cloud platforms and development environments, with MiniMax, “continuously expanding the global AI open-source ecosystem.”

2. The market’s response, in the same session. MiniMax’s Hong Kong-listed shares (MiniMax-W) rose more than 16 percent on 2026-10-23; Zhipu rose more than 8 percent; the China Securities Hong Kong-Connect information-technology index closed up 2.21 percent after touching more than 3 percent intraday. The trading commentary attributes the move to two items in the same news flow: the M3.1 Flash open-weight release, and a policy item — the 2026-10-21 publication by the CPC Central Committee General Office and State Council General Office of the 《关于推动科技创新和产业创新深度融合的意见》 (Opinions on Promoting the Deep Integration of Technological and Industrial Innovation), which the brokerage coverage read as reaffirming integrated circuits and domestic compute as core directions.

3. Two items of context published alongside, which change how the release reads. The same market commentary records two facts worth carrying, because they place the release inside a policy moment rather than a product cycle:

  • On 2026-10-22, the Council of Chairpersons of the 14th NPC Standing Committee decided that the 25th session of the Standing Committee will convene in Beijing from 2026-11-02 to 2026-11-06, and one of the agenda items submitted by the State Council is 审议关于提请审议批准《关于成立世界人工智能合作组织的协定》的议案 — deliberation of the motion to ratify the agreement establishing the World AI Cooperation Organization (世界人工智能合作组织). That agreement was signed in Shanghai in July 2026, with the organisation’s headquarters to be located in Shanghai. Ratification by the NPC Standing Committee is the step that converts a signed international agreement into binding domestic legal effect.
  • On 2026-10-21, the two Offices’ Opinions on integrating technological and industrial innovation were published, with the AI-plus (人工智能+) action and industrial-chain self-reliance running through them.

Institutional Analysis & Commentary:

The interesting fact about the M3.1 Flash release is not its size. It is its timing relative to a filing.

Announcement No. 48’s thresholds are 2 trillion parameters, 10²⁶ FLOPs, or globally leading capability. M3.1 Flash, at 428B total parameters, is far below the parameter threshold — and this is the single most important thing to say about it, because the obvious misreading of this week is that MiniMax published a frontier model and the regime waved it through. It did not have to wave. The release sits under the threshold that would have made the exemption load-bearing, which means it cannot be cited either as proof that the exemption works at the frontier or as proof that it does not. What it can be cited for is something narrower and, in my reading, more consequential: it shows what a regulated firm’s release cadence looks like once a registration path exists. A firm that has just completed pre-export registration formalities for its overseas licensing on 2026-10-22 and releases open weights on 2026-10-23 has demonstrated that the two activities are not in tension in its own operating model. Whether the second was permitted by the first, or simply smaller than the first needed to cover, is not disclosed — and the absence of disclosure is the normal condition of a filing regime.

The market move is the part of this week that should be read as an institutional signal rather than as a price. A 16 percent re-rating on the day of an open-weight release, in a stock that had fallen roughly 88 percent from its March peak and below its IPO price, is not investors valuing 428 billion parameters. It is investors repricing a risk that had been hanging over the sector for seven days: the risk that Announcement No. 48 meant Chinese open-weight releases would stop. The sequence is what matters — draft on 10-16, state-media report that two firms completed filings on 10-22 evening, benchmark-tracker open-weight release on 10-23, double-digit rally on 10-23 — and it describes a market that had been pricing regulatory ambiguity, and repriced when the ambiguity resolved in the direction of “publish and file.” If that reading is right, then the most economically significant document of the past fortnight was not the draft. It was a four-paragraph state-media account of two administrative formalities.

Which raises the harder question, and it is a question about what prices actually discipline here. A share price that rises on a permission is a price that reflects the state’s disposition rather than the firm’s capability. This is the precise mechanism by which an administered channel reallocates risk without allocating rights: the firm’s valuation comes to depend on a filing being accepted, and the filing’s acceptance is discretionary in form even when routine in practice. Nothing in the draft promises that the next filing will be treated the same way. The exemption’s economic value to a listed company is therefore not the licence — the licence is unchanged — but the predictability of the channel, and predictability granted by administrative practice rather than by text is a form of capital that can be withdrawn by administrative practice. Ostrom’s point about rules that survive is precisely that rules made by the same party who applies them, with no appeal route, are not rules but decisions. A market that pays 16 percent for a decision has correctly identified where the value is, and has also correctly identified where it can be taken away.

There is a genuine achievement in this release that deserves to be stated without hedging, because it is the kind that this ecosystem produces too rarely. The named ecosystem list is not decoration. vLLM and SGLang are independent, globally governed inference frameworks whose maintainers owe MiniMax nothing; FlagOS is BAAI’s cross-chip stack; Hugging Face and ModelScope are the two distribution layers. A release that lands in vLLM and SGLang on day one is a release whose portability has been engineered, not merely permitted — and portability, in the strict sense this series has used the word, means the recipient can run the model without the originator. MiniMax has not published which licence governs the weights in the sources reviewed here, and that omission matters; but the framework co-announcement is the more informative signal about intent, because a firm that wanted a component rather than a commons would not spend engineering on making its competitor’s serving stack support it first.

The counterweight in the numbers, and it is not small. The 100-trillion-token figure over 72 hours and the 70 percent cost reduction are service metrics, published by the firm, about its own traffic. They describe an operator’s economics, not a community’s. The relevant comparison is the one this series drew three days ago about DeepSeek: a licence allocates rights in an artefact and nothing about capacity, latency or operational competence; a provisioning record allocates nothing about durability. M3.1 Flash is at present a released artefact plus a firm’s traffic report. What it will become depends on whether anyone other than MiniMax serves it, tunes it, or forks it.

The 2026-11-02 ratification matters here in a way that is easy to miss, and it is the reason these two stories sit together. The World AI Cooperation Organization, headquartered in Shanghai, will be the standing intergovernmental venue for exactly the kind of outward technical cooperation that Announcement No. 48’s facilitation clause names — international standards participation, joint R&D using overseas resources, overseas deployment. An organisation whose founding agreement is ratified by China’s legislature in the first week of November, in the same fortnight as a comment period on an export-control catalogue closing on the 15th, is not a coincidence of scheduling; it is the two halves of one posture. The Catalogue says what may leave. The Organization is the venue in which what leaves is then discussed with the recipients. Read that way, the exemption for published technology stops looking like a loophole and starts looking like the mechanism that keeps the outflow legitimate: you cannot run a cooperation organisation on prohibited transfers, but you can run one on published artefacts.

The checkable question. Two instruments decide whether this was a release or a precedent, and both are public. First: the licence. MiniMax has not, in the coverage reviewed, stated the licence governing the M3.1 Flash weights. If it is Apache 2.0 or MIT, then a firm that had nine days earlier completed pre-export registration released a 428B model into the global commons unconditionally, and the exemption has an operator. If it is a restricted-use or commercial licence, then MiniMax has done what Moonshot did with Kimi K3 — published weights under terms that are not open source in the OSI sense — and the release belongs to a different category, one in which the words “open source” describe a distribution channel rather than a grant. Second: the registrations. Whether DeepSeek’s and MiniMax’s filings are cited again — in the 2026-11 final Catalogue, in MOFCOM guidance, or in the next firm’s announcement — will determine whether 2026-10-22 was the day a procedure became a path, or the day two firms received a courtesy. The procedure is the achievement. A courtesy is not.

Sources:

📡 WeChat Monitor — Forty-Eight Members, Three Tiers, and the Framework the Standards Body Chose to Promote by Name: Reading CAICT’s Agent Open Source Community

1. The instrument. On 2026-10-23, the WeChat account of the China Academy of Information and Communications Technology (CAICT, 中国信息通信研究院) published “智能体开源社区成员单位持续征集中！” — the Agent Open Source Community is continuously recruiting member units. The community was initiated by CAICT, relying on the Cloud Computing Open Source Industry Association (云计算开源产业联盟, CCSIOA), jointly with industry parties, with the stated purpose of “building a collaborative and innovative open-source ecosystem, co-constructing an open, secure and trustworthy agent technology base, and accelerating the enablement of agent technology across every industry.”

2. Membership, by the numbers, from the announcement’s own annex. The community was formed in April 2026. A first cohort of 32 member units covers “basic telecom operators, AI technology enterprises, agent technology enterprises, cloud service providers, internet platform enterprises and network-security enterprises.” In July 2026, a second cohort of 16 member units was formally released at the Agent Internet Forum of the 2026 China Internet Conference, bringing the total to 48. The full list, as published, distributes the field across categories with unusual clarity:

  • Telecom operators and their cloud arms: China Telecom, China Unicom, Tianyi Cloud, China Mobile (Suzhou) Software, China Telecom AI.
  • Platform and AI companies: Tencent, Huawei, Baidu, Kingsoft Cloud, ZTE, China Electronics Cloud.
  • Security vendors: Venustech (启明星辰), Knownsec (知道创宇), Antiy (安天), Asiainfo, Riverse (瑞数), Anpro (悬镜安全), Chengdu Yuzhian.
  • Certification and research: Tai’er Certification Centre (泰尔认证中心) — CAICT’s own certification body — the Institute of Software, Chinese Academy of Sciences, and Beijing Institute of Technology (Zhuhai).
  • One foreign-facing foundation: The Open Agentic AI Foundation (OAAIF).
  • Ecosystem and hosting: Modelers Community (魔乐社区); a long tail of provincial digital-government, data and integration companies.

3. The governance design, stated in three tiers. The community runs on a “council plus special working groups plus expert team” (议事会+专项工作组+专家团队) three-tier collaborative model. The council (议事会) is the “daily decision-making body, responsible for deliberating and deciding major community matters.” The working groups are the “execution units of daily work” and are split into an industry open-source application working group and a technology open-source application working group, focused respectively on sector-scenario deployment and technical-system construction. The expert team “provides guidance and suggestions on the community’s technical development direction.”

4. The five mandates, verbatim in substance — and the sentence that matters most. The community’s stated mission is to solve “common pain points in agent technology R&D, standards construction, security governance and commercial deployment,” across five dimensions:

  • (i) Core technology base co-construction — focused on agent open-source technology application, and here the announcement names its object: “推广 OpenClaw 等开源智能体框架的开源应用” — to promote the open-source application of open agent frameworks such as OpenClaw, to push shared open interaction protocols, and to lower the barrier to industry access.
  • (ii) Trustworthy application standards co-research — joint industry-academia-research work specifying norms around protocol interaction, interface adaptation, security and trustworthiness, and establishing a standards-landing and testing-verification system.
  • (iii) Industry ecosystem trends — regular publication of research reports and forward-looking topics to support members’ technology-route and strategic decisions.
  • (iv) Innovative deployment scenarios — a supply-demand matching platform offering “full-lifecycle incubation and promotion services from proof of concept to commercial deployment.”
  • (v) Community public services — technical exchange, thematic seminars and closed-door roundtables (闭门座谈会), drawing on “industry think tanks and expert resources.”

5. The recruitment mechanics. Applicants download a member-unit application form by QR code and email it to two caict.ac.cn addresses in the open-source and software security department of the Cloud Computing and Digitalisation Institute. The third intake closes 2026-11-06 (Friday) 17:00.

Institutional Analysis & Commentary:

Three things in this document are new relative to what this series has already documented about CAICT’s apparatus, and the third is the one that changes a judgement.

First, the governance form is a council, not a TOC, and that distinction is doing structural work. In the foundation model this series has repeatedly used as the benchmark — OpenAtom, Mulan, the Linux Foundation — the decisive body is a technical steering committee whose authority derives from having reviewed and merged code, and whose members are typically maintainers. A council (议事会) is a different animal: it is a decision-making body for “major community matters,” which in practice means budget, membership, work-plan and direction. The working groups execute; the expert team advises. No tier in this structure is constituted by the act of contributing code. That is not a criticism of the design — a community convened by a standards institute to reduce adoption friction across 48 organisations reasonably organises around alignment rather than patches, and councils are how alignment is achieved. But it does mean the community’s decisions are made by representatives of member organisations, which is a stakeholder body, not a maintainer body. The two produce different things: stakeholders produce consensus documents, maintainers produce merged code.

Second, and this is the sharpest available evidence of a pattern this series has been documenting for months: the standards body is promoting a specific third-party framework by name. “推广 OpenClaw 等开源智能体框架” — promote the open-source application of open agent frameworks such as OpenClaw — sits in the community’s first mandate, before standards, before research, before scenarios. In a document whose other four mandates are all verbs of co-construction, co-research and co-creation, mandate (i) is the only one with a named product. Now weigh this against what the same institute has been saying about openness: SJ/T 12304-2026 grades model openness on an L1–L5 / O1–O4 scale; the OSCAR apparatus issues governance-maturity and transparency gradings; the preceding weeks’ documents have been careful to describe rather than endorse. A named framework in a mandate is an endorsement, and it is a stronger instrument than any grade, because a grade ranks an existing thing while a mandate directs future effort. The practical consequence is that 48 organisations — telecom operators, provincial digital-government builders, security vendors — now receive an official signal about which agent framework to build their scenario deployments on. That is not how a commons grows. It is how a standard gets adopted, and it works.

Third — and this is the observation that should temper the two above — the membership list contains a genuinely unusual entry, and it is worth asking what it is for. The Open Agentic AI Foundation (OAAIF) is a member of a Chinese state-institute-convened community. So is the Institute of Software at the Chinese Academy of Sciences, and Tai’er Certification Centre — CAICT’s own certification arm, which sits in the membership as a unit alongside the enterprises it may later certify. The presence of the certifier inside the community it may assess is the structural feature a reader should hold onto: when the same organisation’s certification body is a member of the community whose practices it will later evaluate, the assessment is not third-party, and the standards co-researched in mandate (ii) are being written by parties who will be graded against them. That is efficient — it eliminates the adversarial step — and it eliminates the adversarial step.

Compared with the OSCAR conference programme this series read two days ago, this document is a materially different kind of instrument, and the difference is instructive. OSCAR’s six items were all apparatus: results released, standards launched, groups founded, reports compiled, practices recognised. Nothing in it could be falsified. The agent community announcement contains one falsifiable claim and one countable figure: 48 member units, named in an annex, with a third intake closing on a stated date. Membership is a fact about an organisation’s choices, published, checkable, and it can grow or stall. That is a weaker instrument than a standard and a stronger one than a conference, because a register of who joined is a record of who committed. The right way to read the number 48 is not as evidence of success but as a baseline — the thing against which the third intake’s outcome, published or not, will be measured.

The honest counterweight, stated because it is real. Multi-stakeholder communities convened by a neutral technical institute across operators, vendors, security firms and academia are a genuine public good in a fragmented field, and this one is addressing a real coordination cost: agent development in China in 2026 is split across dozens of incompatible tool-calling conventions, and a shared protocol layer would save every integrator money. Ostrom’s design principles do not require a maintainer body; they require clearly defined boundaries, congruence between rules and local conditions, and — critically — that the people bound by the rules have a say in modifying them. On that last principle this structure is silent, as every CAICT instrument this series has read is silent: the announcement says who deliberates and who executes and who advises, and says nothing about how a member contests a council decision, or how a working group’s output can be rejected by the members it binds.

The checkable question. Two observables, both public within a year. First: does the third intake’s outcome get published? The first two cohorts were announced with full named lists. If the November intake is reported with names and a new total, the register is a real instrument and the community is growing on a measured trajectory. If intakes thereafter are reported only as counts, or not at all, then the list of 48 is a founding photograph rather than a membership. Second, and more decisive: what does OpenClaw’s position look like in twelve months — is it named in the standards the working groups produce, or is it named only in the mandate? A framework that is promoted in a mandate and then specified in a standard has become the de facto reference implementation for a national ecosystem, whatever its licence says. A framework named in a mandate and absent from the standards was advertising. The difference between the two is whether the community’s technical working groups write about OpenClaw or write for it, and that is a document anyone can read.

Sources:

🔍 Commentary — Three Days, Three Documents, and the Reclassification of the Permission Step

On 2026-10-22 the state’s own broadcaster reported that two of the country’s leading AI firms had completed the paperwork and open-sourced.

On 2026-10-23 one of those firms released 428 billion parameters of weights and its shares rose sixteen percent.

On 2026-10-23 the country’s principal telecommunications research institute published the names of forty-eight organisations that had joined its agent open-source community and opened a third intake.

None of the three mentions the other two. Read in sequence, they describe one thing: the permission step in Chinese open source has been relocated, and it is now being operated deliberately rather than applied suspiciously.

Start with what a licence does and has always done, because the relocation is legible only against it. Open source’s founding institutional move was to remove the permission step. A licence is a standing grant to strangers: no application, no membership, no fee, no nationality, no one to ask. Foundations, technical steering committees and maintainer authority grew up afterwards, and their job was to keep the permission step from being reinserted — which is why they are constituted by contribution rather than by representation, and why a patch from a stranger is a peer’s patch. Everything this series has examined in the past month — graded openness, maturity tiers, transparency levels, standards-launch conferences, membership registers — has been an insertion of a permission step at some layer. The puzzle was how to describe that without overstating it. This week supplies the missing half of the description: the permission step has not been reinserted at the layer where the code lives. It has been installed at the layer where the release is announced.

That distinction is not a loophole. It is a design, and it has a name in the literature on administrative instruments: the front-end register. A regime that must approve an infinite variety of private decisions faces an information problem it cannot solve from a distance. A regime that requires every decision to be declared before it is executed converts the information problem into a filing problem, at negligible cost, and retains the option to intervene only where the filing says something worth intervening in. Announcement No. 48’s registration triggers are all declaration points — ownership transfer, outbound licensing, change of controller, and whatever MOFCOM later designates. DeepSeek and MiniMax declared. Their declarations were accepted. The releases continued. A system that accepts nearly every declaration and keeps every declaration is not a gate; it is a census with a veto it has not used. And a census is a far more consequential instrument than a gate to anyone who has to plan, because a gate tells you no and a census tells the state everything.

Now the second half, which is where the judgement must be careful, because it is easy to slide from description into a story about strategy. Reading this week as a coordinated choreography — draft, register, release, ration — would be a satisfying narrative and it would outrun the evidence. What the evidence supports is narrower and, I think, more durable: three separate institutional mechanisms were built by three separate bodies, in the same fortnight, and they happen to be mutually reinforcing. CAICT built a membership community for agents. MOFCOM and MOST built a registration census for technology exports. A firm and its market priced the relief. No document reviewed here shows one of these bodies coordinating with another, and the sequence of dates is tight enough to be scheduling and loose enough to be coincidence. The correct claim is not conspiracy. It is institutional complementarity: in a system where the same handful of organisations populate the standards institutes, the industry associations, the certification bodies and the ministry-adjacent research organs, the instruments they build will fit together, because the people building them share a model of how the field should be organised. That is not a criticism. It is how every administered field works, including in the jurisdictions that describe themselves as market-led.

Which brings the analysis to what is genuinely at stake, and it is not whether China is “really” open source. That question has been asked, in this form, for a decade, and it has no answer because it is not a question about a fact. The answerable question is the one this week puts sharply: when a release requires a filing, what does the licence still allocate? And the answer is: everything it allocated before, plus less. The MIT or Apache grant is unchanged, and unchangeable — the weights that are downloaded are permanent, forkable, and unforgeable. But the decision to release now has a procedural cost, an administrative counterparty, and a timing dependency. A commons produced by actors whose every publication passes through a register is a commons with a release schedule that is not set entirely by its contributors. That is a real institutional difference from a commons whose releases are set by whoever has the commit bit, and it does not disappear because the register is lenient.

The counterweight, and it is not a small one, is the thing this week actually demonstrates and that a Western reader is likely to miss entirely. In the same season, one jurisdiction refused a model release pending a classified review and took it back nine days later; another jurisdiction published its thresholds in numbers, wrote its exemptions into the operative text, opened a comment period with an email address, and then had its head of government’s broadcaster explain at length what was not covered. Whatever else is true, the second set of actions is inspectable and the first is not. Predictability is a public good that costs something to supply, and it has been supplied here. A firm deciding where to publish in 2027 will notice, and the fact that it will notice is the reason this policy exists.

The last observation is about the forty-eight. A membership community of 48 named organisations, convened by a standards institute, promoting a named framework, with a council that decides and working groups that execute, is the most complete expression of one institutional form this series has documented: open source as a coalition to be assembled rather than as a commons to be maintained. It will produce standards, test regimes, research reports and pilot deployments, and it will do so on a schedule, because that is what a council with 48 members and a work plan does. What it will not produce spontaneously is the thing that makes a commons durable — a maintainer in another country, on their own hardware, whose patch gets merged because it is correct, and who has no relationship to anyone in the room.

The checkable question. One number, published by someone, within twelve months, will settle the character of all three mechanisms: how many of this fortnight’s releases were made by parties who are not in any of the registers? If the next significant Chinese open-weight release comes from a lab that is not named for development in a COPU outline, not among the 48, and not on the export register, then the registers are describing an ecosystem they do not control and open source is still doing its job. If every significant release now comes from a registered party, filed in advance, announced into a community whose council sets the agenda, then the commons has acquired a front office — and the question that will follow, for everyone using Chinese weights outside China, is not whether the licence is permissive. It will be who is on the other side of the filing desk when the next one is declined.

Sources: