Open-Source

On 2026-09-18 Beijing’s Municipal Bureau of Economy and Information Technology and its Development and Reform Commission issued the Action Plan for Accelerating the Development of the Token Economy (2026-2028) — six dimensions, ten measures, first provincial-level token-economy policy in the country. Its two most institutional proposals are not subsidies: a graded evaluation standard for ’token factories’ scored on model-adaptation count, token throughput, first-token latency, cache hit rate and PUE; and a token distribution platform to be written into the new-infrastructure plan, offering developers one interface, model aggregation and metered settlement. Set beside that, two financing schemes already in operation: Beijing E-Town’s 2-billion-yuan token loans closed in mid-September through six banks, and Haizhu district’s guarantee-innovation product from August — both using token production and consumption as a credit signal for firms with high R&D and no collateral. And a consumer layer nobody planned: token packs sold like mobile data, token credit-card rewards, tokens handed out at dumpling restaurants, tokens resold on Xianyu, against a State Data Administration series that went from 100 billion daily tokens in early 2024 to 500 trillion by May 2026. The economics being assembled here is older than it looks: 词元 — word-plus-currency — was adopted as the official term in March 2026, and the last Chinese government to make a unit of consumption into a unit of account ran into a problem that Mises described in 1912 and that no rediscount window ever solved.

A first endogenous answer to the question yesterday’s file left open. On 2026-10-24 a foreign-ministry progress report said open cooperation requires ’toolchains, test environments and related support that remain to be built.’ Three National AI Application Pilot-Test Bases now publish enough to be read: the automobile base published on 2026-10-23 as a China Mobile-led industry AI foundation claiming 112.7 EFLOPS of managed compute under the stated position ’neutral, does not build cars, capabilities fully open’; the household-appliance base launched in Qingdao on 2026-10-27 with Haier Smart Home as 承建方 and Huawei, China Mobile and Alibaba Cloud as first ecosystem partners, running an open-call (揭榜挂帅) procurement architecture for common problems; and the culture-and-tourism campus, opened in Hangzhou on 2026-09-27, with five named models, four named datasets covering 40,000-plus ancient buildings and 400,000-plus components, and three-year targets of 33 scenarios and 15 standards. And on 2026-11-01, one day before the Wuzhen summit opens under the theme ‘Open Source, Openness, Co-construction, Shared Benefits,’ the country’s first open-source-themed experience hall opens — a three-floor exhibition of, among other things, the pilot-test bases. The institutional question is not whether the test environments get built. They are being built. It is what a test environment is when a state builds it: a factory for the missing middle between paper and repository, assigned to existing organisations so that somebody bears the residual — and, by design, not a commons.

On 2026-10-24, at the 2026 Global Governance Forum in Xiongan, Foreign Minister Wang Yi released the two-year progress report of China’s AI Capacity Building Inclusive Plan, issued by the Global AI Innovation Governance Centre on the Ministry of Foreign Affairs domain. Read end to end it is not a technical document. Its ten actions carry ten numbered cases, and exactly one of them is an open-source artefact — FlagOS, a multi-chip system software stack supporting 21 chip vendors and 1,454 operators. The other nine are platforms, deployments, seminars, toolkits and statements, culminating in WAICO, signed at Shanghai on 2026-07-16 by 29 states and now holding 38 founding members, described in the report as the first intergovernmental organisation in the AI field. The document’s fourth lesson states the thesis plainly — openness is the precondition for inclusive development — and then, in the same paragraph, says what is missing: toolchains, test environments and related support that remain to be built. Verified against the report’s own inventory, and set beside two documents from the same fortnight at other registers: a MOFCOM think tank arguing in China Daily on 2026-10-13 that permissive licences and stable release cadence, not platform-building, are what produced Qwen’s 150,000-plus derivatives; and the 2026-10-20 State Council Information Office briefing where MIIT reported the 10-billion-download figure that this series first read on 2026-10-06 as 100 billion. That discrepancy is not a correction but an instruction in how these numbers are made — and it is the reason the checkable question below is about repositories, not figures.

The Central Government Procurement Center’s 2026 desktop operating system framework agreement closed on 2026-07-01 with three products listed — UnionTech UOS V20, Zhongke Fangde V3.1, Kylin V10 — and HarmonyOS absent. The reason is item 1 of the MOF/MIIT Operating System Government Procurement Requirements Standard (2023 edition), verified for this briefing in the PDF from gks.mof.gov.cn: a starred, non-scoring indicator requiring that the operating system be same-source compatible with ARM, LoongArch, MIPS, SW64 and x86 CPUs. MOF’s own transmittal notice makes starred indicators substantive requirements that cannot be awarded points, only met or failed. HarmonyOS V1.0 was, in January 2026, the first and still the only desktop operating system certified at Security-Reliability Level II — and it still failed the first page of the form. A WeChat practitioner essay published on 2026-10-28 supplies the explanation from the vendor side: four walls — core architecture, ecosystem, procurement rules, commercial control — and the decisive one is that a distro built on Huawei’s core would turn an independent systems vendor into a customisation subcontractor whose API cadence is set by a competitor. The institutional point is precise: a rule that scores can be met by improving the artefact; a rule that is starred and non-scoring can only be met by changing the artefact’s architecture or the vendor’s identity. China’s procurement form has made the second kind a gate, and a supplier that has just won the safety contest is the case that reveals what such a gate is actually selecting for.

On 2026-10-03, the Day of German Unity, Aleph Alpha released Kolibri — 78.1B parameters, Apache 2.0, and, on its own account, trained from scratch in Germany and Finland with no base-model dependencies. Then page 53 of the 189-page technical report names the models that wrote the reasoning traces: GLM-5.2 and GLM-5.3 from Z.ai, and Qwen3.8-27B from Alibaba. Section 3.1.1.5 concedes the point in the company’s own words — ‘we acknowledge that several of our teacher models were built in China, so they carry known political biases on sensitive topics.’ Twenty-five days earlier the same company had published a study measuring exactly that: six Chinese models returned a balanced answer to 17–41 percent of 967 politically sensitive prompts. Verified for this briefing: the study publishes no score for Kolibri itself, the model card carries a warning that Kolibri ‘may reproduce political biases present in its training data,’ and a third-party reading has found that the study tested no model from Z.ai — the lab behind two of the three teachers. The open-weight regime’s defining property is not that weights are free. It is that the chain of custody is written down — because a licence cannot be negotiated in a way that only works if a ministry asks first.

On 2026-09-16, fourteen minutes apart, a Hugging Face account called ChaosGPT pushed 71 byte-identical files into two of the most downloaded official model repositories in the world: Qwen/Qwen3.8-27B as Pull Request #198, and deepseek-ai/DeepSeek-V4.1-Flash as commit db668511. Neither company merged anything. The payload was an Italian-language autonomous attack-agent toolkit — model abliteration, exploit-dataset poisoning, jailbreak Modelfiles, an infinite-loop agent called smart_chaos.py — with no model weights, no tokenizer, no training data, and no reference to the target models at all; every base model it names is a Llama. Antiy CERT named the form ’latent pollution’ (潜伏式污染) and dated its report 2026-09-21. Verified for this briefing: the Qwen pull request is now gone (HTTP 410, ‘This pull request was deleted’), while the DeepSeek commit db668511 is still reachable and still serves smart_chaos.py — 131 files against 62 on main. The attack’s real target was not either model. It was the merge request itself — the one artefact in a model repository that carries a company’s credibility while carrying none of its review. And in the same week, the largest open-source forum on the English-speaking internet banned AI projects outright, and its moderator wrote down, in public, the reason the ban can work: ‘There’s no objective way to implement this policy. The rule is written so that, from a moderation perspective, we’re afforded discretion.’ Three institutions, three answers to the same question — who is allowed to attach their artefact to your name.

On 2026-09-07 the Supreme People’s Court issued the country’s first national judicial rule-making document on AI disputes — 24 articles, Five Parts, effective as authoritative guidance — and Article 13, on open-source software, is the first time a Chinese court of last resort has written a liability safe-harbour for open-source developers and providers into a general rule. The conditions are narrow and worth reading precisely: the court weighs the type of open-source licence, the specific content of the rights restrictions, the security and compliance measures taken, and the degree of disclosure; and a developer or provider who supplies code modules free of charge and publicly explains their functionality and security risks may be found not liable when a third party’s use causes infringement. Eleven days earlier, on 2026-10-25, a state-informatization industry outlet published an account of why units have begun halting blind domestic-system replacement, and named the mechanism precisely: the acceptance criterion has changed, 国产化率 is now a reference item and no longer a hard threshold, and the binding test is whether the system carries the business, whether faults can be cleared quickly, whether migrated data stays intact, and whether efficiency visibly declines. And on 2026-10-22 the Wall Street Journal reported that DeepSeek raised USD 15 billion in a second private round, nearly double its target, and is accelerating toward a STAR Market listing in Shanghai. Three documents, three different instruments — a court rule, an assessment standard, and a cap table — all deciding the same thing: what counts as a defensible act.

On 2026-10-23 Moonshot AI released Kimi K3.1 Pro to its own API, its apps and its enterprise customers — and announced that the weights will follow on 2026-11-04. The model is 2.8 trillion total parameters, activates 104 billion, and carries a one-million-token context window; by the parameter count it is the largest open-weight model in the world. Three days earlier, CCF’s Open Source Development Technical Committee convened a seminar in Chengdu at which its chair, Academician Wang Huaimin of NUDT, judged that Chinese open source has entered the ‘participating and integrating’ stage, that open source is a strategic fulcrum of AI innovation and technological competition, and that the road to ‘global open-source leadership’ still runs through bottlenecks in platform-grade projects, international communities and long-term mechanisms. In a separate and much smaller item, a Reddit count of the twenty-three projects that smartphones and browsers actually run on found that eleven of them had been maintained over the past year by one or two people — xz by one, at 97 percent of its changes, with no confirmed funding since the 2024 backdoor. Read together, these three items locate the binding constraint: at the top of the parameter curve weights are becoming conditional artefacts published by the largest firms on their own schedules, and at the bottom of the stack the software nobody can do without is held by one person per project.

On the evening of 2026-10-22, Yuyuantantian — the new-media arm of CCTV — reported that DeepSeek and MiniMax had both 顺利履行登记手续, successfully discharged the pre-export registration formalities, and completed 开源和海外授权使用, open sourcing and overseas licensing. Seven days earlier, Announcement No. 48’s exemption for technology ‘already in the public domain by means of open sourcing’ had been a clause on paper; as of that evening it is a clause with two named firms’ registration records behind it and a documented outcome. The next morning, 2026-10-23, MiniMax released the weights of M3.1 Flash — 428B total parameters, 23B active — on Hugging Face and ModelScope, and its Hong Kong shares rose more than 16 percent while Zhipu gained over 8. Read together with the two documents from the same fortnight — a 32-plus-16-member agent open-source community being recruited by CAICT for a third intake closing 2026-11-06, and the NPC Standing Committee’s 2026-11-02 agenda item on approving the agreement establishing the World AI Cooperation Organization — three events in three days describe a single institutional arrangement: the permission step has moved off the licence and onto the release form, and the state has begun to operate it as a channel rather than as a gate.