A lab that open-sources its best model says it will stop

Where things stand (8 Nov 2026)

Nothing qualifying has been published since the window opened on 1 Nov, and no resolver ruling exists on this question.

Who is eligible. I re-downloaded Epoch's ECI file (https://epoch.ai/data/eci_scores.csv, 8 Nov; its values have shifted slightly since 1 Nov) and checked Hugging Face through its API on 8 Nov.

• DeepSeek. V4.1 Flash 1017 (17 Oct, MIT licence) has ECI 159.31, DeepSeek's highest.
• Moonshot. Kimi K3 (ECI 158.51, open). The K3.1 Pro weights were committed to HF on 2 Nov and announced on 4 Nov, so Moonshot stays eligible.
• Zhipu. GLM-5.3 (ECI 156.91), weights on HF since 25 Aug.
• Alibaba. Qwen3.8-2.4T-A95B went on HF on 8 Aug. Its model card says it is the base of Qwen3.8-Max and "for the first time… brings a Qwen-Max-class model to open release". That keeps Alibaba eligible until about Aug 2027.
• Not eligible. Meta's Muse Spark Open (155.42) sits 6.6 points below Muse Wonder (162.02), and Meta did not call it its flagship. OpenAI, Google, xAI, Anthropic and ByteDance have no open flagship.

New this week

• Moonshot. The K3.1 Pro model card does not repeat Moonshot's 23 Oct "open-weight principles" (no open weights that "extend the absolute frontier of cyber capability"). That text predates the window anyway. It also would not qualify, because Moonshot says K3.1 Pro is below that frontier.
• Alibaba. Qwen4 is not out. Alibaba posted a teaser on 6 Nov. Slides show Max as the flagship and 27B as the open-weight model, and Qwen 3.7 already went closed without explanation. A closed Max with no stated reason is a silent switch, which the criteria exclude. Manifold's "Qwen 4 open weights by 25 Nov" market has fallen to 61%.
• World Internet Conference, Wuzhen (3 Nov). The theme was "open source", and the messages were strongly pro-open:
  • Premier Li: "encourage open source… [but] the safety reins must be held tighter… oppose a digital iron curtain".
  • Joe Tsai (Alibaba): "only open source can drive widespread adoption".
  • DeepSeek's Gao Xi: open source is "the inevitable choice". He admitted it amplifies risk and said DeepSeek is working on grading its models.
  • Sources: Sina/The Paper (http://finance.sina.com.cn/jjxw/2026-11-03/doc-inixnpzn9808529.shtml); SCMP (https://www.scmp.com/tech/big-tech/article/3373262/deepseek-engineer-defends-push-open-source-ai-smarter-humans-despite-higher-risks, 4 Nov); BigGo (https://finance.biggo.com/news/5c11d48f-aed3-47e1-91de-f1685f54cccb).
• Rules. Comments on MOFCOM's draft export-control catalogue close on 15 Nov. The draft gives open-source releases automatic licences and exempts API access (China Briefing, https://www.china-briefing.com/news/china-technology-export-controls/, 3 Nov).
• Longer-run rule-making. The NDRC said on 3 Nov that comprehensive AI legislation will be accelerated (Caixin, 4 Nov). TC260 is drafting a national standard on open-source model security.
• Zhipu. It co-wrote a "Frontier Open-Weight AI Risk Management Framework" with red and yellow lines (Concordia AI, https://concordia-ai.com/research/frontier-open-weight-ai-risk-management-framework/, 21 Sep). Zhipu has hinted that GLM-5.5 is coming, possibly with 2.5–3.5T parameters.

Reference class

Labs that once open-sourced their flagship and later held it back: OpenAI (GPT-2), Meta, Mistral, xAI, Alibaba (Qwen-Max, 3.7, Omni-Flash), Zhipu (GLM-4), Baichuan, 01.AI.

• Almost all of these closures were silent or commercial. The one clear safety-cited stop is GPT-2 in 2019, from a safety-focused lab.
• That is roughly 1 in about 28 lab-years of open-flagship status, and 0 in about 15 Chinese lab-years.
• About 20 eligible lab-years remain to 2031, which would naively give 30–45%. I discount this because Beijing champions open source and Chinese firms seldom attribute decisions to regulators or call their own models dangerous.

Paths to YES by 2031 (judgemental)

• Binding Chinese rules plus a lab statement (≈12%). A final catalogue, AI law or TC260 standard bars open release of top-tier models or confines them to China (≈27%), and a lab announces it is complying, for example in a model card or HKEX filing (≈45%). The listed companies are Zhipu, Alibaba, and Moonshot if its IPO goes ahead in early 2027.
• Voluntary safety withholding (≈9%). Zhipu under its framework, or Moonshot under its cyber ceiling, says the next flagship's weights will not ship.
• Commercial closure that also cites safety or compliance (≈3.5%).
• A non-Chinese lab becomes eligible and then stops (≈1.5%).
• Edge cases (≈2%).

Combined, these come to about 25%. The hazard series ends a little higher, at 27%, which allows for faster capability gains.

Hazard series

• Q4 2026: 0.9%. Possible triggers are Qwen4, GLM-5.5 and the final MOFCOM rule, but the signals favour openness.
• 2027–28: 1.7–2.0% a quarter. Rules are implemented, Moonshot's prospectus may appear, and open models reach today's restricted cyber levels.
• 2029–31: falling from 1.6% to 0.9%. Labs that close silently lose eligibility.

Cumulative probability, computed in Python: 8.3% by end-2027, 15.3% by end-2028, 20.4% by end-2029, 24.1% by end-2030, 27% by end-2031. The 1 Nov run had 26% by 2031. Wuzhen lowers my near-term number; Zhipu's framework and the legislative and standards work raise the later years slightly.

Strongest case against

• Lower (10–15%). Beijing treats open weights as strategic, so any tiering would hit only "core" models. Labs then comply silently, as Alibaba did with Qwen 3.7.
• Higher (40%+). Capabilities rise quickly, models are already near-frontier in cyber, Beijing's tiered-governance blueprint includes domestic-only release, and listed labs have to disclose changes of strategy.

Markets are thin and only loosely related. Manifold gives 23% that China restricts open release above a threshold by end-2026, 79% by end-2027, and 66% that the US or China bans such releases by 2030. Those markets count licensing requirements and need no statement from a lab. My forecast moves with the forecast that an open-weight model comes within 2 points of the best on Epoch's index (≈30% by 2031).

Reading adopted

These count: a non-temporary decision not to release the current flagship's weights, or to release only smaller or domestic-only versions, that cites safety, security or rules. Delays, filtered releases, "for now" pauses and ceilings above the current model do not.

Question. By the last day of each quarter, has a covered developer that released open weights of its most capable model in the preceding 12 months publicly said it will stop releasing its most capable models' weights, citing safety, security, or government rules?

Definitions.

• Released open weights of its most capable model: in the 12 months before the statement, the developer made publicly downloadable, without individual approval and under terms allowing at least research use, the weights of a model that, on the day of release, either (a) scored within 2.0 points of the highest-scoring model the developer had released by then on Epoch AI's general Epoch Capabilities Index (one data vintage), or (b) was described by the developer at release as its most capable or flagship model. Leaked weights do not count.
• Says it will stop: states a policy or decision that it will not, or will generally or by default not, release publicly the weights of its most capable models, or of all its models above a stated capability or compute level that its current most capable model exceeds. Saying it will release weights only of smaller or weaker models counts, and so does saying it will make them available only inside its home country.
• Citing: the statement gives as a reason safety, misuse, security (including national security or theft), or a law, regulation, export control, or official government request or guidance. Commercial or competitive reasons alone do not count, but a statement giving both kinds of reason does.
• Publicly said: on the developer's own domain or official social-media accounts, in a developer-authored document such as a model card, licence or policy, or in a filing with a regulator or stock exchange. As in Public confirmation (A2), executive interviews, employees' personal posts, leaks and anonymous-source reporting do not count.

Resolves YES if a qualifying statement is published within the Resolution window. The developer need not have stopped yet. A later reversal does not undo a YES.

Resolves NO / Exclusions. Statements by developers not eligible when they speak; temporary pauses, safety-testing delays and licence changes that still allow download; silent switches to closed weights; other parties' statements about a developer's plans.

Resolving source. The developer's own publications and filings defined above, with Epoch AI's ECI data at epoch.ai/eci and developers' release pages or official repositories to establish eligibility. Apply the Resolution window and Resolution date (A3) conventions.