⚠️ 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 — 2026-11-04

1. Five artefacts, one week, four different licences — and none of them a discussion. Everything below is checkable against primary files, and all of it was first published on or before 2026-11-03.

ArtefactWhat it isLicenceWhat the licence adds beyond permission to use
zai-org/ZCodeCoding-agent harness, desktop + web + CLI + agent runtime, 6,973 files / ~1.03 M lines at publicationApache-2.0Nothing. OSI-approved, unconditional except notice retention.
zai-org/GLM-5Documentation and skills repo for the GLM-5.x family (README, tech report, skills/, benchmark images)Apache-2.0Nothing.
zai-org/GLM-5.3-Flash (Hugging Face)WeightsMITNothing.
zai-org/GLM-5.3 (Hugging Face, ModelScope)141 safetensors shards, 152 files“GLM-5.3 License”, license: otherAbove USD 10 bn trailing-12-month revenue, an MaaS operator must pass “Z.AI’s security review”, whose “scope and method shall be reasonably determined by Z.AI.” Also: use “must comply with applicable laws and regulations.”
moonshotai/Kimi-K3 (Hugging Face)2.8 T-parameter weights + training/inference code“Kimi K3 License”Over USD 20 M in any 12 consecutive months, an MaaS operator needs “a separate agreement with Moonshot AI”; over 100 M MAU or USD 20 M monthly revenue, the product must display “Kimi K3” “prominently.”

Both custom licences are drafted as MIT plus conditions: an unrestricted grant of rights to use, copy, modify, merge, publish, distribute, sublicense, sell, run, deploy and fine-tune — and then numbered paragraphs that attach consequences to scale. They are near-models of each other, which is not surprising: they were written in the same jurisdiction, in the same year, by firms solving the same problem.

2. The calibration is the interesting part, and it is too precise to be accidental. Moonshot’s threshold is USD 20 M a year; Z.AI’s is USD 10 bn a year — five hundred times higher, and roughly the size of a single hyperscaler’s AI line item rather than of any actual startup. A condition that no plausible licensee will trip is not a restriction; it is a reserved right. What it buys the grantor is not revenue but standing: the terms exist, they are written down, they are enforceable in principle, and they cost nothing in adoption. The 100 M-MAU attribution clause does the same work on the brand axis: it converts a right nobody reserves into a name that must appear on someone else’s product. What is not reserved is revocation — neither licence claims a right to withdraw the grant from users who already have the weights, and that is the term that would matter most if it were missing. It is worth saying plainly, because it is the reason these documents are licences rather than terms of service.

3. The one condition that is not a reserved right, and the one that points outward. “The Licensee’s use of the Software must comply with applicable laws and regulations” appears in both custom licences and is a clause worth reading literally rather than politically. It is jurisdiction-indeterminate by construction: it imports whatever law applies to the licensee, wherever the licensee sits, at the moment of use. A European SME running GLM-5.3 locally is therefore bound by a term whose content depends on European law; a US cloud operator by US law; and a joint venture by whichever law a court later says. Compare the OSI-approved licences on the same table: Apache-2.0 and MIT bind the licensee only to the notice and to the warranty disclaimer. The difference is not that one regime restricts and the other does not; it is that one regime’s restrictions are fixed text and the other’s are a pointer. A founder-licence that points at “applicable laws” converts the licence into an interface to a legal system — which is a strictly weaker commitment than a licence whose obligations are legible from the file itself.

4. Reported downstream traffic, stated at the strength the sources support. At Wuzhen on 2026-11-03, Alibaba chairman Joe Tsai gave the day’s numbers: 470-plus Qwen models open-sourced, 4 billion-plus cumulative downloads, 330,000-plus derivatives, and ModelScope at 33 million users across 239 countries and regions, carrying 280,000 models, 40,000 datasets and 12,000 MCP entries. These are company figures and the first three are not independently verifiable in a way that separates model downloads from file fetches; they are recorded as reported. CCTV’s commentary account Yuyuantantian, relayed by China News Service the same day, added a third-party series: per OpenRouter, weekly token calls to Chinese large models had exceeded the United States’ for 27 consecutive weeks, with four of the global top five weekly models Chinese and DeepSeek first. OpenRouter is route-level data and it measures calls, not downloads — a distinction that matters below.

Institutional Analysis & Commentary:

Start with the thing that is genuinely better than it was a year ago, because the record deserves it.

In September 2026, after a paying subscriber published packet captures of the ZCode client shipping his workspace to Alibaba Cloud, Zhipu AI did the expensive thing: it apologised, removed the pipeline, and on 2026-09-20 published the harness under Apache-2.0 — desktop, web, backend, agent CLI and agent runtime, 6,973 files, 1,033,262 lines, one commit, feat: open source. A firm that responds to a disclosure by publishing source has chosen the harder form of accountability when a press release would have been cheaper. Whatever follows in this briefing should not obscure that.

Now the part that the licences cannot fix, and it is the part that decides whether a grant is an institution or an announcement.

Apache-2.0 is a property rule: it tells every person who holds a copy exactly what they may do with it, and it is enforced by the copyright interest the licensor gives away. It says nothing about who decides what goes into the next copy. That is a governance question, and governance is not in the licence — it is in the repository’s plumbing. On that plumbing, the record as of 2026-11-04 is:

  • Contributors: three. MBearo (3 contributions), zRzRzRzRzRzRzR (1), and an anonymous ZCode Source Sync Bot (4). No aggregate list of outside contributors exists because there are none.
  • Issues: off. has_issues: false. The repository’s “Issues” tab does not exist in the sense it exists on any community project; GitHub’s open_issues_count of 11 counts pull requests, and there are no pull requests either.
  • Pull requests: zero, ever. GitHub search across the repository returns 0 pull requests in any state and 0 issues. Not “few.” Not “moderation queue.” Zero.
  • Commits: flattened, and continuously flattened. Eight commits since 2026-09-20: an empty Initial commit, the 1.03 M-line feat: open source, two hand-authored updates, and then four commits by the sync bot, each carrying the same trailer — Source-Commit: <sha> Source-Sync: zcode-open-source-v1. The tree moves; the history does not.
  • What “sync” means in tree terms. Reconstructing each tree from the GitHub API: 2026-09-20 → 6,973 blobs / 72 MB; 2026-09-24 → 7,060 / 73 MB; 2026-10-10 → 16,880 / 187 MB; 2026-10-19 → 18,238 / 201 MB; 2026-10-20 → 18,266 / 201 MB; 2026-10-30 → 17,203 / 190 MB; 2026-11-03 → 17,237 / 190 MB. One commit on 2026-10-30 reported 828,056 lines changed — 216,552 added, 611,504 deleted — under the message chore: sync ZCode source, with no further description.

The diagnostic is not that the repository is run this way; it is what the tree can and cannot answer. A working note from Z.ai’s own provider-refactor effort is still public in the repository (docs/working-memory/provider-refactor/steps/todo-34-model-selection-async-state-and-repo-wiki-draft.md), dated 2026-08-28, before the disclosure. It shows the vendor running an internal task list with a real defect, real log evidence and a real fix. So the maintenance process is not missing. The maintenance process is internal, and the publication channel is an export of its output. That is a coherent way to run a product. It is also, exactly and completely, the thing that licences cannot touch: an unconditional grant over a snapshot, renewed at a bot’s convenience, at a time the vendor chooses, with a diff nobody outside the vendor can review page by page.

Which brings the day’s two candidate institutional readings into contact, and the contact is the argument.

Reading one, and it is generous: Chinese AI labs have built a publication institution — high-quality weights, permissive licences, global hosting, real downloads — and the governance keeps up as the projects mature. Reading two: what has been built is publication without custodianship, and the download numbers measure the first while the contributor counts measure the absence of the second.

The evidence does not split neatly. On the weights side, publication is complete in the sense that matters legally: a 141-shard checkpoint on two platforms, a licence, a chat template, and — this is the part that shows care — a Hugging Face repository with fifteen commits of its own, dated 2026-08-28 through 2026-10-14, adding harness notes, community evaluation results and chat-template fixes. That is a maintained artefact. On the software side, where a project has to accept other people’s code to be an institution rather than a product, the record reads differently: zcode-plugins, the marketplace repository that is ZCode’s actual extension point, was created on 2026-08-26 and holds 102 pull requests; GitHub reports 0 merged by the merge API. Its commits are authored by a bot handle — “ZCode Marketplace Bo…” — and titled github-pr-1:, github-pr-45: — the pull request’s content has been imported as a commit, so the pull request itself sits open by design, or because nothing closes it. Either reading is worth saying out loud; the README does not. Against that, three external pull requests have been merged in GLM-5’s history (#56, #166, #157) and those are real, small and documentation-only. Set them beside 17 open issues on GLM-5, of which the most recent on the public record are agent-quality complaints: quotas consumed by a defective orchestrator, a low-effort reasoning run that “falsely certified a defective” result, an agent task whose root cause was asserted without a control. These are exactly the reports a maintainer needs, and they are arriving at the place where a maintainer is supposed to be. The question the record cannot settle is whether a maintainer is there.

Williamson is again the correct instrument, and here it gives a precise answer to a vague worry. L2 is the formal rule: the licence. At L2, the grant is real and it is strong — Apache-2.0 on the harness, MIT on the small model, custom terms on the frontier weights that almost nobody will trip. L3 is governance: who decides, who merges, who is answerable. At L3, the decision rights are entirely inside the firm. Open source’s actual innovation was never the licence alone; it was the discovery that you can separate the copyright grant from the decision right — give away the first completely, and distribute the second through a governance process that anyone with a patch can, eventually, influence. Here the first is given away and the second is retained, in full, by a four-contributor repository with no issue tracker. That is not hypocrisy and it is not fake open source. It is a real grant attached to a real product, and the two have different owners in a way the licence does not disclose.

And this is where the ZCode case earns its place in an institutional series rather than a security one. The event is well documented elsewhere: v3.12.3 packaged workspaces including .git — 86.6 percent of a 313 MB archive was git objects, LFS and reflogs — encrypted the payload with a key wrapped to a server-held RSA public key, and POSTed it to Alibaba Cloud OSS, triggered before prompts and on task completion; a small public-repo snapshot was accepted by the server; CAICT confirmed the bucket held no data and NSFOCUS confirmed bucket and objects deleted; the full audit reports promised in September had not, on the public record, appeared. The institutional point is narrower and it is about publication as a remedy: publishing a snapshot can restore verifiability of the current tree; it cannot restore auditability of history that was never published. The public repository can answer “does the client send my code today?” — and the researcher who first reported the flaw answered it himself, finding the snapshot pipeline gone. It cannot answer “when was this added, who approved it, and why was it on by default?”, because the file that would answer that was replaced, not reverted. A licence over a snapshot is a property rule over a present state; accountability for a past state needs history, and history is a governance artefact, not a licence one.

Keep the comparison honest by turning it on the reference system too. Corporate open source everywhere publishes with a licence and curates internally, and nobody demands that every vendor’s release be a community decision. The distinction that survives scrutiny is not “corporate versus community.” It is whether a stranger’s patch has any path to the next release. In GLM-5, on the documentation axis, a stranger’s patch did — #166, itzKLAUS, merged 2026-10-15, one file, skills/glm-master-skill/SKILL.md. In the weights, no stranger’s patch is conceivable, and there the right instrument is not a contribution process but a feedback loop, which is precisely what the Hugging Face repository’s fifteen commits of external evaluation results are. In ZCode and zcode-plugins — the surfaces where an ecosystem would form — the loop is an import: pull requests land as bot commits, or do not land. Two of these three arrangements work. Only one of them is called open source on the slide.

Which returns the argument to Wuzhen, and to the thing a theme cannot do. The summit’s own materials, in state media’s own framing, have open source delivering downloads, affordability, capacity for the Global South, and a model for international cooperation. Every one of those claims is supported by the index and unsupported by the ledger. Downloads happen at Hugging Face and ModelScope; the terms are set by the vendor; the maintenance is a bot; the decisions are a roster. That is a defensible industrial arrangement — it lowers the cost of world-class weights to zero, which is a genuine and large public benefit, and the briefing has said so before. But it is not the arrangement that produces the thing the theme says it is producing, because an ecosystem is not a download count. An ecosystem is the set of people who will keep a project alive after the company that made it stops caring, and that set is legible in exactly one place: the list of people with merge rights who do not work for the company. On the record assembled this week, that list is short and mostly empty, and it is not a secret — it is a contributors endpoint.

The checkable questions, all cheap, none rhetorical.

(i) Does any of these repositories accept a pull request from an outsider in the next sixty days? If zcode-plugins’ design — review-in-the-PR, merge-as-import — is a deliberate marketplace model, say so in the README, and the zero-merged figure stops being evidence of anything. If it is not deliberate, it is the exact point at which “we open-sourced it” and “it is open source” diverge.

(ii) Will the frontier licence thresholds ever bind, and will anyone publish the security review it entitles the vendor to demand? A USD 10 bn trigger is a reserved right; if a review is ever conducted under it, its existence and its criteria are the test of whether the clause is a real rule or a placeholder. The USD 20 M threshold in Kimi’s licence is low enough to bind on real MaaS operators — and “a separate agreement” on undefined terms is a club, not a public licence. If the first such agreement is signed, its terms matter more than the weights.

(iii) Do the internal working notes keep appearing in the public tree? The presence of docs/working-memory/ in an exported repository is either an oversight or an argument: it lets an outsider see that a real review process exists and that outsiders are not part of it. Both readings are useful. Only one of them makes the project an institution.

(iv) Does any aggregate contributor number from a lab survive being split into “employees” and “not”? The sector’s headline metric is downloads, and downloads cannot answer this. The answer takes one API query and no cooperation — which is the point: an institution can be inspected without the institution’s permission. That property is what open source licenses are for, and it is the property most at risk when the licence is unconditional and the plumbing is closed.


🔍 Commentary — The Licence Is a Property Rule; the Ecosystem Is a Governance Question

Three findings from the record above, stated as claims rather than as worries.

First: the licences being written in China right now are calibrated to be tripped by nobody, and that is their design. MIT plus a USD 20 M MaaS condition, MIT plus a USD 10 bn condition, MIT plus a 100 M-MAU attribution requirement — each is an unrestricted grant with a reserved right sized to the grantor’s actual exposure. That is intelligent commercial drafting and a genuine improvement on “research-only” weights. It is not, however, a contribution to the licence commons: a term whose only function is to preserve an option creates no obligation anyone can build on, and the OSI-compatible core — Apache-2.0 for the harness, MIT for the small model — is doing all the work of openness while the custom text does all the work of governance.

Second: the jurisdiction pointer is the clause to watch. “Use must comply with applicable laws and regulations” is the only line in either custom licence that binds a normal user, and its content is whatever law reaches that user. Over a decade this is how a permissive grant becomes conditional without the text changing: the file stays the same and the world around it acquires new rules. A project whose obligations are defined by reference is a project whose obligations are defined elsewhere.

Third, and the one that matters most for this series’ subject: the download is becoming the unit of account, and the download is not a governance event. ModelScope carries 280,000 models and 40,000 datasets for 33 million users; Qwen reportedly sits at 4 billion cumulative downloads and 330,000 derivatives; GLM-5.3 alone shows 2.36 million downloads on Hugging Face and GLM-5.3-Flash 6.3 million; and OpenRouter reports Chinese models ahead on weekly tokens for 27 straight weeks. Not one of those figures distinguishes a user who ran inference from a person who will file a bug, and not one of them rises when a stranger’s patch is merged. A system that measures adoption and cannot measure custodianship will optimise adoption. That is not a moral failing; it is what measurement does. The corrective is not a lecture about community — it is a second number beside the first, published by the same people who publish the first, and it costs one API call.

The Chinese labs have demonstrated, this year, that they can ship frontier weights under licences the world can use, at a speed and price that has visibly changed the global market. The open question — the one the Wuzhen theme cannot answer and the repository can — is whether any of it will still be maintained by anyone outside the originating firm in 2029. A licence guarantees that you may fork. Only a governance arrangement produces anyone who will.


Sources: zai-org/ZCode — repository, Apache-2.0, commits, tree sizes via GitHub API (read 2026-11-04) · zai-org/ZCode — commit history (8 commits, 2026-09-20 → 2026-11-03) · zai-org/ZCode — NOTICE.md (data-egress disclosure) · zai-org/ZCode — internal working note todo-34-…repo-wiki-draft.md, dated 2026-08-28 · zai-org/zcode-plugins — 102 PRs, 0 merged by API (read 2026-11-04) · zai-org/GLM-5 — repository, 63 commits, 17 open issues, Apache-2.0 · zai-org/GLM-5 — PR #166, merged 2026-10-15 · Hugging Face — zai-org/GLM-5.3 (GLM-5.3 License, 2,356,195 downloads) · GLM-5.3 License — full text, USD 10 bn security-review clause · Hugging Face — zai-org/GLM-5.3-Flash (MIT, 6,316,631 downloads) · Hugging Face — moonshotai/Kimi-K3 model card · Kimi K3 License — full text, USD 20 M MaaS clause · InfoQ — 智谱把 ZCode 开源了,然后呢? (2026-09-23) · 东方财富/央广财经 — 智谱股价三个交易日累计跌逾16%,公司回应 ZCode 数据上传争议 (2026-09-23) · Bodega One — ZCode tried to upload your Git history. What actually happened. (updated 2026-10-13) · CellCog — Z.ai Open-Sources ZCode After Silent Git Uploads: The Record (2026-09-21) · Sina Finance/澎湃 — 蔡崇信乌镇谈AI:中国开源模型下载量、衍生模型数等关键指标保持领先 (2026-11-03) · China News Service — 乌镇峰会观察:中国开源AI大模型持续"圈粉" (2026-11-03) · Eastmoney/广州日报 — 阿里巴巴蔡崇信乌镇谈AI:只有开源开放 才能普惠普及 (2026-11-03) · Dataconomy — Alibaba’s Tsai calls open source AI’s only sustainable path (2026-11-03) · CAC — 2026年世界互联网大会乌镇峰会部长级会议召开 (2026-11-03)