⚠️ 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 — 2026-11-12
🏛️ Policy & Regulation — The Safe Harbor Gets a Track Record: Automatic Licensing, the Four-Day Window, and What Registration Actually Schedules
1. Where the instrument stands, as of this morning. Announcement No. 48 of 2026 — the joint MOFCOM/MOST draft revision of the Catalogue of Technologies Prohibited or Restricted from Export — was published 2026-10-16 and opened for public comment until 2026-11-15. That is three days from this briefing. The interim pre-export registration requirement, by contrast, took effect on the day of publication and continues until the revised catalogue is formally issued and enters into force, with no linkage to the consultation deadline. Registration is live; the catalogue is not. Anyone reading this in Beijing has a deadline on Sunday and an obligation that does not end on Sunday.
2. What changed in the ministry’s own words since the draft. This series covered the announcement on 2026-10-17 and noted the open-source carve-out. What that entry could not yet report is that MOFCOM released a spokesperson’s Q&A alongside the announcement that goes considerably further than the catalogue text, and this is where the institutional content sits. Asked what consideration went into the 23 AI technologies, the spokesperson answered with a policy of positive facilitation, and the sentence is worth reproducing exactly:
“To support the building of an open-source innovation ecosystem for artificial intelligence, for technology exports in cases including the open-sourcing of AI technologies, Chinese enterprises building and operating their own compute centres abroad, going abroad for chip tape-out and manufacturing, conducting joint R&D using overseas resources, and participating in international standard-setting, facilitation measures such as automatic licensing apply, basically achieving ‘report-and-go’ (即报即出); for technologies that have already entered the public domain through open-sourcing, publication of academic papers and similar means, it is proposed not to include them within the scope of export administration, creating a favourable environment for international exchange and cooperation in AI technology.”
Three things are being said at once here, and they are not the same thing. (i) A safe harbor: technology already public is not administered at all — it exits the regime rather than being licensed within it. (ii) A fast lane: technology that is not yet public but is being released, or is being deployed through channels the state considers low-risk, is licensed automatically — the ministry’s own gloss is that it is essentially notification. (iii) A judgement test: for models in the restricted category, overseas localised deployment is an export, while API access, app/web interaction, SaaS and lightweight on-device deployment are not — because the model keeps running at home and only the result crosses. MOFCOM stated that of the AI technology in scope, things are “precise and limited,” and that after implementation “the vast majority of AI technology may still be freely exported.”
3. The registration queue has begun producing case law, and it is the part that matters. On the evening of 2026-10-22, CCTV’s affiliated outlet Yuyuantan (玉渊谭天) disclosed that DeepSeek (深度求索) and MiniMax (稀宇科技) had completed the relevant registration procedures and walked their large models through the compliance process for open-source release and overseas authorisation. On 2026-10-23 MiniMax published its own announcement confirming it had completed technology-export registration for “several recently released and expected-to-be-released models,” covering both open-source release and overseas licensing. Yuyuantan added that overseas users were calling for more Chinese open-source LLMs to follow, and that a clear, predictable compliance channel would help steady global users’ confidence. A brokerage analyst quoted by STAR Market Daily drew the inference the companies would want drawn: “with core sensitive technology strictly controlled, the overseas path for general-purpose open-source large models has instead become clearer.” A founder of an overseas-facing B2B AI startup said that before the new rules, SMEs had no clear boundary and “many overseas businesses were pushed conservatively to avoid risk” — and that the two registrations now give the industry a reference template.
Institutional Analysis & Commentary:
A safe harbor written in a draft is a promise. A safe harbor with a completed first filing is a precedent, and the difference between them is the entire story of this fortnight.
Start with what the design actually is, because it is more interesting than “China restricts model exports,” which is how it was largely reported abroad. The regime does not control publication; it controls transfer. That single distinction explains every exemption in the document. A model’s weights on a public URL are not a transfer of technology from a domestic holder to a foreign recipient — they are already outside. An API call returns a result while the artefact stays home. A developer downloading from a repository is receiving a public thing. A company licensing weights to a foreign cloud is transferring restricted technology, and that is what gets a file. MOFCOM’s formula is precise: the question is whether restricted technology moves from inside China to outside it. On that criterion the exemptions are not loopholes; they are the definition applied consistently. An institutional economist should say plainly that this is a better-drafted control than most: it names its unit of account (the technology artefact, not the revenue, not the user), it states where the line falls, and it publishes the reasoning.
Then the hard part, which the first filings have now made testable. Note what Yuyuantan is reported to have stressed: “the exemption looks at the actual state of the technology, not the label the company affixes” — a company cannot wrap controlled core technology as “open source” or “a paper” and walk it out. And a related point from the compliance commentary: a plan to open-source something later does not establish the exemption today. Both sentences are doing the same institutional work, and it is the work this series flagged on 2026-11-06: the carve-out converts is this technology already public? from a question a regulator investigates into a representation by the releasing firm. What the fortnight has added is that there is now a queue that produces a record. DeepSeek and MiniMax did not merely assert; they filed, and a state outlet and then the company itself said so. The representation has acquired an audit trail, which is the minimum condition for a representation to function as an institution rather than as an honour system. That is a genuine improvement and it should be recorded as one.
Third, and this is the observation that the framing “encouraging open source” obscures: the fast lane schedules the release, and a scheduled release is a governed release. Automatic licensing for “the open-sourcing of AI technologies” is, read coldly, a procedural right to publish — report and go. That is a strong commitment, and by the standards of technology-export regimes anywhere it is unusually liberal. But an automatic licence is still a licence. It presupposes that publishing frontier AI is an act the state has jurisdiction over and has chosen to permit generically rather than specifically. The alternative design — the one that treats publication as outside the regime altogether unless downstream transfer is restricted — was available, was arguably the more natural reading of the “already public” principle, and was not chosen. The difference is not rhetorical. It determines whether a lab’s release calendar is a matter of internal decision or of filing timing, and it determines what happens on the day a lab wants to publish something the fast lane has not yet been defined to cover — a genuinely open question, since “automatic licensing” is described in the commentary as a mechanism that “does not exist under the current regulatory framework” and is “expected to be further defined in future rules or guidelines.”
Fourth, the part that should interest anyone who cares about open source as a commons rather than as a distribution channel. Look at what the exemptions optimise for. Open-sourcing, self-operated overseas compute, overseas tape-out, joint R&D using foreign resources, international standards participation — every one of these is a channel through which a Chinese firm’s capability reaches a foreign user. The regime’s generosity is distributed along the outbound axis: the more ways a domestic actor has to get its technology into the world, the more facilitation it receives. What is conspicuously absent from the list is anything on the inbound axis — a foreign developer’s ability to participate in the upstream project, a maintainer’s standing to review, an outside contributor’s claim on the governance. That asymmetry is the same one this series has measured for six weeks, and it is now written into a trade regulation. The instrument makes it easy to ship a model and silent on what happens when someone ships back a patch.
Fifth, the honest counterweight, because the record supports one. The compliance commentary is right that a vague boundary is itself a cost: firms that cannot tell whether an API call is an export will over-comply, and the SMEs it quotes say exactly that — they were “pushed conservatively.” Writing the exemptions in detail is the cure for that, and it is a public good. A regime that publishes “here is what is not controlled” is doing something a regime that publishes only “here is what is controlled” cannot. And the fact that a state outlet and a company both made the first filing public within seven days suggests the queue is not a black box. Set against the version of this regime that many observers expected in July 2026 — curbs on overseas access to China’s top models — the published outcome is materially more open. That should be said without hedging.
The synthesis, and it is uncomfortable for both camps. The people who read Announcement No. 48 as China closing its model ecosystem are wrong: the text, the Q&A and the first two filings all point the other way, and a general-purpose open-source model now has a clearer path abroad than it did in September. The people who read it as China endorsing open source are also wrong, and more subtly: the state has not endorsed open source as a mode of production. It has identified open-weight publication as the lowest-friction channel for a national capability to travel, and built a fast lane for the channel. A commons that is valuable because it is a commons and a commons that is valuable because it is a distribution mechanism produce identical press releases and completely different institutions — and the test that separates them is not in this document. It is whether, on the day a foreign contributor’s patch arrives at the upstream project built by a firm that has just enjoyed “report-and-go,” there is anyone on the other end whose job is to answer.
Checkable questions.
(i) Will the final catalogue and its implementing rules publish the automatic-licensing criteria and the turnaround time — and will a filing be refusable? “Automatic” is a claim about administrative behaviour, and it is checkable in exactly one way: publish a service standard. If automatic licensing means “we approve within N days, and refusal requires written reasons,” it is a right. If it means “we process it,” it is discretion with a friendly name.
(ii) How long did the DeepSeek and MiniMax registrations take, and were either of them asked to change a release? The first two filings are the precedent the whole regime will be read against. A same-week turnaround with no modification requests is a different institution from a six-week review with redactions.
(iii) Does any published version of the exemption list contain a contribution clause — that is, anything about inbound participation in the projects being exported? On everything published so far, it does not. That absence is the finding, and it is the one question an open-source reader should insist on asking of a document that uses the words “open-source innovation ecosystem” eleven times.
Sources: MOFCOM via IT之家 — 商务部回应《中国禁止出口限制出口技术目录》拟增列人工智能等技术:确保安全可控,鼓励开源开放、合作共享 (2026-10-16; full spokesperson Q&A; the 即报即出 automatic-licensing sentence; API/SaaS/on-device exclusion) · MOFCOM & MOST — 公告2026年第48号 (official announcement text) · White & Case — China proposes new export controls on AI, biopharma, and photovoltaic technologies (2026-10-26; 14 prohibited / 26 restricted; interim registration; automatic licensing described as new to the framework) · Baker McKenzie — China Proposes Expanded Export Controls on AI, Biopharma and Solar Energy Technologies (technical thresholds: 10²⁶ FLOPs, 2 trillion parameters, top-10 benchmarks, 10,000-accelerator clusters) · Geopolitechs — Announcement No. 48 full translation including Annex 1 (2026-10-16) · Regulations.ai — China tech export consultation closes on 15 November (2026-11-11; registration outlives the consultation; a plan to open-source later does not establish the exclusion today) · 南方都市报 via 腾讯新闻 — AI出海划红线,DeepSeek、MiniMax完成登记 (2026-10-26) · 科创板日报 via 凤凰科技 — DeepSeek、MiniMax完成AI模型出海登记 (2026-10-25; MiniMax’s 2026-10-23 announcement; analyst and startup-founder quotes) · 驱动中国 — DeepSeek、MiniMax完成开源和海外授权,AI技术出口登记路径走通 (2026-10-23) · 腾讯新闻 — AI管制扩容,API和SaaS却被豁免:卡的是技术出境,不是服务跨境 (2026-10-19; the label-versus-state argument) · This series’ 2026-11-06 briefing — the public-domain carve-out as a representation nobody audits · This series’ 2026-10-17 briefing — the original Announcement No. 48 entry
🏗️ Institutional Change — The Helper Robot as a Donated Upstream: AgiBot Gives Lingqu OS to OpenAtom, and OpenLET Donates Data Instead of Selling It
1. The two events, in order. On 2026-10-29, at the 2026 AgiBot Innovation Day in Shanghai, AgiBot (智元机器人) — through partner and co-president Xiong Yan (熊彦) — released Lingqu OS 1.0 (灵渠OS 1.0), described as an embodied-intelligence-native operating system covering the full stack from hardware enablement to application support, validated across its production robot line. The company announced it would fully open-source Lingqu OS 1.0 by end-2026, contributing 200 million lines of code validated in mass production, 5,000+ ecosystem software packages, an end-to-end build toolchain and pre-integrated embodied-AI components. Simultaneously AgiBot announced it had become a platinum donor of the OpenAtom Open Source Foundation and donated the Lingqu OS project to the foundation, with the two parties to co-build a Chinese humanoid-robot open-source community spanning body and hardware, data and evaluation tooling, operating system, and the “large brain / small brain” model layers — plus joint development, test validation, and embodied-AI talent cultivation.
2. The foundation-side event, five days later. On 2026-11-03, the OpenLET Embodied Intelligence Developer Conference was held in Shijingshan, Beijing, under the guidance of the OpenAtom Foundation. OpenLET — the foundation-guided embodied-intelligence community, seven months old — reported: 70+ co-building organisations, over 10,000 developers, over 100,000 minutes of real-robot open data, and more than 7 million total downloads. Its working committee chair is Leng Xiaokun (冷晓琨), chairman of Leju Robotics (乐聚机器人). From an automotive production-line deployment with partner Zhaofeng Mechanical & Electrical — 15 “Kuafu” robots running routinely, 98% loading success, helping produce over 60,000 components — the community reported that the resulting high-quality real-robot data, the Kuavo-VLA industrial domain model and the motion-control toolchain had been formally contributed to OpenLET, with an industrial in-the-loop dataset donated at the same time. OpenAtom chairman Xie Shaofeng (谢少锋) framed the conference as bringing together “policy, scenario, technology, capital and entrepreneurial services — five in one” to help teams convert technology into deliverable capability.
3. The five signings, which show where the boundary work is happening. Shijingshan district opened Shougang Park as a live training ground and launched a humanoid-robot traversal competition; China Automation (中国自控), a central SOE, opened flexible automotive assembly and smart-warehouse scenarios; China Mobile Research Institute, Zhongguancun Ubiquitous Networks Institute, MediaTek and Leju signed on embodied-AI device–network coordination and quality assurance; eight universities including Tsinghua and Peking University formed an OpenLET Beijing University Developer Alliance; and an embodied-intelligence incubator plus a Physical Intelligence OPC were launched.
Institutional Analysis & Commentary:
Two donations happened here, and only one of them is the kind this series has spent six weeks describing as missing. The difference is the unit of account, and it is worth being precise about.
Take the Lingqu OS donation first, and give it its due. A production-validated operating system is a real thing to give away, and the specifics matter: an OS that abstracts 15 device classes and 100+ peripherals, with deterministic end-to-end latency jitter under 4 microseconds at 1,000 Hz task load, is exactly the layer that every robot company would otherwise rebuild. Handing it to a neutral foundation with a platinum-donor commitment and a named governance counterpart is the structurally correct move, and it is the opposite of the pattern this series documented on 2026-10-26 in the Alibaba postmortem — where projects without neutral-foundation backing (Rax, easyexcel) died and those inside CNCF/Apache/OpenJS lived. AgiBot is doing the thing that survives.
But notice the transaction form. The gift is of a codebase at a point in time, plus a commitment to continue. Nothing in the reported arrangement creates an obligation on AgiBot to fund review of outside contributions, to seat an outside maintainer on the project’s technical governance, or to answer a stranger’s patch within any period. The foundation supplies legitimacy and administration; it does not supply labour. And the code being donated is code that AgiBot has already written for its own production line — the marginal cost of the gift is near zero, and the strategic return (a standard others build on, a hiring pool, a moat at the integration layer) is substantial. That is not cynicism; it is the correct economics of a platform donation, and it is why platform donations are relatively easy to obtain. Williamson would place this at L3: a governance form chosen because it minimises the transaction cost of getting other firms to build on your substrate. The commons is the instrument, not the objective. Easy gifts are gifts of things already paid for.
Now the other donation, the one that is genuinely new, and the reason this entry exists. An industrial in-the-loop dataset. The OpenLET report says the data, the domain model and the toolchain from a live automotive deployment were contributed to the community, and that an industrial in-the-loop dataset was donated alongside. Sit with what that is. The scarce input in embodied intelligence is not code; it is real-robot trajectories collected in production conditions, with the failure modes included, and the interviewee says the quiet part out loud: “real deployment-state data is the most valuable data for us — that data is itself a very high moat.” A moat, contributed to a shared pool, by a listed company, in a jurisdiction where the prior six weeks of this series have been about instruments that fund deployment and never maintenance. This is the first item in that run of briefings in which the donated asset is neither a platform nor a certificate but a costly, non-reproducible input to everyone else’s work. Ostrom’s second design principle — congruence between rules and local conditions — has an unstated corollary that matters here: a commons works when the thing pooled is the thing that is scarce. Code was never scarce in Chinese open source. Production data is.
Third, the honest limits, because the pattern is still the pattern. (i) The dataset, the model and the toolchain came from one deployment with one partner; a community whose “in-the-loop data” is a single automotive line is a seed corpus, not an institution. The number that would settle it — how many distinct contributing firms, and how many of them are not Leju’s integration partners — was not reported. (ii) The community is guided by the foundation and chaired by one company’s chairman, which is the same structural tension as the Xi’an certification centre co-housed with the winning supplier: legitimate, disclosed, normal for an industry consortium, and not independent in the way a standards body is. (iii) The conferences’ governing resource is stated as “policy, scenario, technology, capital and entrepreneurial services — five in one”, which is a description of an industrial-policy delivery channel, not of a maintainers’ forum. The five-in-one formula is the tell: it tells you what the instrument is for. (iv) And the accumulated community metrics — 10,000 developers, 7 million downloads, 100,000 minutes — are production metrics, not reciprocity metrics. They measure consumption of the commons; none of them measures contribution to it.
Fourth, the connection to the first entry, because the two events are the same fortnight for a reason. Both are about what a firm does with an asset that has already paid for itself. Announcement No. 48’s fast lane exists because a firm that has finished training a model can release it at near-zero marginal cost and gain distribution; the OpenLET dataset exists because a firm that has finished a deployment can donate its logs at near-zero marginal cost and gain an ecosystem. In both cases the incentive to contribute is strongest precisely where the asset is already sunk — which is the best possible news for the volume of open contributions in China and the worst possible news for their location. Code that is already shipped and data that is already collected flow outward readily. The work that has not been done yet — the unglamorous, expensive, non-sunk work of reviewing someone else’s patch on a Thursday — is the work no donation in either story covers.
Fifth, the counter-observation that keeps this honest. It would be easy and wrong to conclude that nothing durable is being built. OpenLET is seven months old with 70 organisations and eight universities in a formal alliance, and it has a stated pipeline from pilot to contributed corpus — which is more institutional continuity than most Chinese open-source communities of the past decade achieved in five years. And Lingqu OS going to a foundation rather than being held in-house is a governance choice, made before it had to be, which is exactly the choice that makes a project survive its founding team. The prediction this entry is willing to make is narrow and falsifiable: if a dataset donated by a company that is not Leju’s partner appears in OpenLET’s corpus within twelve months, the community has become an institution. If the corpus stays inside one supply chain, it is a consortium with a foundation’s letterhead.
Checkable questions.
(i) When Lingqu OS 1.0 is actually released at end-2026, what licence, and what does the project’s governance document say about who approves a change? “Donated to the foundation” and “open to outside maintainers” are different claims. The licence will be public on day one; the maintainer list will tell you which claim is true.
(ii) How many distinct firms have contributed data to OpenLET, and what share of the 100,000+ minutes comes from organisations other than Leju and its integration partners? The community is publishing a total. A distribution would convert a metric into an institution.
(iii) Does the donated industrial dataset retain the failures — the unsuccessful grasps, the aborted cycles, the near misses — or only the passes? A dataset of successes teaches a model what works; a dataset of failures teaches a field what to fix. The “98% loading success” figure in the same press release is either the beginning of the answer or the evasion of it.
Sources: 新浪科技 — 智元发布具身智能原生操作系统-灵渠OS1.0,将开源2亿行代码、5000+软件包 (2026-10-29; Lingqu OS 1.0 specifications; platinum donor status; donation of the project to OpenAtom; end-2026 full open-source commitment) · 科创板日报 via 腾讯新闻 — 智元与开放原子开源基金会共建中国人形机器人开源社区 (2026-10-29; the four cooperation directions: body and hardware, data and evaluation tooling, OS, large/small brain) · 机器人前瞻 — 专访智元熊彦:让技术走出一家公司的围墙 (2026-10-30; open-source rationale; AIMA ecosystem 80,000+ developers, 5.6 million dataset downloads; 元苼计划 ¥300M invested in 2026, ¥2B over five years, 255 universities, 511 enterprises, 40 partners receiving deployment incentives; “real deployment-state data is the moat”) · 环球网 via 新浪财经 — 开放场景、共享技术、协同创新!开发者生态激活具身智能发展新动能 (2026-11-04; OpenLET: 7 months, 70+ co-builders, 10,000+ developers, 100,000+ minutes of real-robot data, 7 million downloads; 15 Kuafu robots at 98% loading success across 60,000+ components; data/model/toolchain contributed to the community; industrial in-the-loop dataset donated; the five signings; Xie Shaofeng’s “five in one”) · 经济观察报 — OpenLET具身智能开发者大会在京举办 (2026-11-03) · This series’ 2026-10-26 briefing — the company-donation pattern and the three survival indicators
🔍 Commentary — Two Kinds of Gift
First, today’s two entries are the same transaction performed on different assets, and naming the transaction is the whole of the analysis. In Beijing, a ministry builds a fast lane so that finished technology can leave the country with a filing. In Shanghai and Shijingshan, a robotics company gives away a finished operating system to a foundation and a finished deployment’s data to a community. In neither case is the gift costly at the margin, and in both cases the gift is exactly what the giver would have wanted to circulate anyway. That is not a moral failing. It is the structure of contribution incentives everywhere, and it is why open source works at all: people give away what they have already built. The question an institutional analyst has to ask is not are these gifts sincere but what does a system built only of such gifts fail to produce.
Second, the answer is the same in both entries, and this series has now arrived at it from six different directions. The 2026-11-09 entry found a foundation with ¥450 million pledged and twelve unanswered issues. The 2026-11-10 entry found a ¥20.7 million state grant restricted to a “cultivation platform” and spent on testing, licensing, training and hosting. The 2026-11-11 entry found a city that built a complete certification-to-procurement loop and no line for upstream review. Today: a donated OS with no maintainer obligation, and a donated dataset whose donation is the easy part. Every instrument in this ecosystem is optimised for the moment an asset crosses a boundary outward — model to the world, code to a foundation, data to a community. None is optimised for the moment an asset arrives inward. Not because anyone decided that; because outbound is where the fundable transaction is. A ministry can facilitate an export and a company can donate a repository, and both fit a contract that somebody was already going to sign. Nobody was already going to sign a contract to review a stranger’s patch.
Third, the correction this week owes its own framing. For a fortnight this series has used “deployment machine with a certification front end” as the description of Chinese open source, and today’s stories look like more of the same. They are not quite. A donated operating system under a foundation is a governance choice that outlives its founder’s enthusiasm, and an industrial dataset in a shared pool is a commons contribution in the strict Ostrom sense — a costly input, made non-excludable, by an actor who could have kept it. Those are real, they are new, and a critique that cannot register them is not an analysis but a posture. The honest position is that the ecosystem has become genuinely good at the outbound half of the commons and has not yet built the inbound half — and that the outbound half is not nothing. It is the half that makes a project exist.
Fourth, the sentence to keep, and it is a prediction about what to watch rather than a verdict. The two structures on display today differ in one respect that matters more than licence or governance: a fast lane and a donation both specify what leaves; neither specifies who answers. So watch the receipt rather than the shipment. The English-language test of whether China’s open-source ecosystem has crossed a threshold is not how much code it releases, how many models it exports, or how many developers a community claims — it is whether, some Thursday in 2027, a patch arrives from someone with no relationship to the founding company, and somebody whose job description includes that queue replies the same week. Nothing in Announcement No. 48 prevents it. Nothing in the Lingqu donation funds it. It would cost less than the certification regime, less than the voucher scheme, and less than one line of the ¥245 million Xi’an contract — and it is the only piece of infrastructure in this entire run of briefings that no budget line currently builds.
Sources: MOFCOM via IT之家 — 商务部回应《目录》修订 (2026-10-16) · 新浪科技 — 智元发布并开源灵渠OS 1.0 (2026-10-29) · 环球网 via 新浪财经 — OpenLET 具身智能开发者大会 (2026-11-04) · This series’ 2026-11-11 briefing — the certification-to-procurement loop and the clause that was never written · This series’ 2026-11-10 briefing — the filings and the restricted platform grant · This series’ 2026-11-09 briefing — the ¥450 million commitment and the register-versus-queue finding