The US requires a licence to publish frontier model weights openly

Where things stand (30 Oct 2026)

Nothing qualifying exists, and nothing that would qualify is pending.

• Federal Register. I queried the API on 30 Oct for rules from the Commerce export-control bureau (BIS) since 25 Aug and for today's public-inspection list. There is no AI model-weights rule. The only AI orders since May are EO 14409 and the October "Super Intelligence" renaming order.
• White House presidential actions page (30 Oct). No new AI order.

The administration has explicitly chosen not to do this.

• EO 14409 §3(c) says nothing in it authorises "a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution" of models. The question excludes this order by name.
• On 4 Aug the White House told labs that open-weight models are exempt from its voluntary pre-release review framework (Yahoo/WSJ/Bloomberg, 5 Aug). The National Cyber Director said a regulatory regime would "strangle growth."
• The Mythos/Fable restrictions in June came through company-specific "is informed" letters (Mayer Brown). The question excludes those.
• A mandatory-review order was scrapped in May, and a national-regulator idea stalled in August, after calls from Musk, Zuckerberg and Huang (tech-insider).
• CNAS (26 Oct) judges that Washington "seems unlikely to limit the diffusion of open-weight models anytime soon" (CNAS).

No bill would qualify. I read the texts via the Congress.gov API:

• The FRONTIER Act (H.R. 9925) offers only court-reviewable emergency orders.
• Warner–Schatz–Kim (S. 5576) and S. 5061 require pre-release access but give no power to block.
• The kill-switch bills (H.R. 9917, S. 5417) leave coverage to later rulemaking.
• S. 2938 is excluded by the question itself.
• Anthropic's draft bill allows only court-ordered suspension.
• The only reported "block power" is in the unpublished Thune–Cruz–Klobuchar draft, which Cruz has since moved against (OrcaRouter).

The pro-open coalition is broad. OpenAI says frontier policy should not "become open-weights policy by another name" (OpenAI). More than 270 firms signed the open-weights letter. Even Anthropic rejects a ban.

Reference class

• US prepublication licensing of software is rare and contested. The main cases are 1990s encryption controls, which were rolled back after Bernstein, and the narrow carve-back for 3D-printed gun files.
• The Biden administration also declined to control published weights (January 2025 diffusion rule).
• Legal hurdles are high: the First Amendment and prior-restraint safeguards, the informational-materials exemption in IEEPA and export law, and the major-questions doctrine (Lawfare, Sept; Lawfare, 29 Oct; Just Security).
• Prior: well under 1% a quarter now, rising if dangerous capabilities spread.

Paths to YES

1. A pre-release approval law covering open releases. Most likely after 2028. I start from the sister forecasts: about 56% for any frontier safety law by 2031 and 45% for required pre-release access. I assume only about 20–35% of such regimes would carry a power to refuse, and about 65% of those would cover open publication.
2. A BIS rule that carves weights above a threshold out of the "published" exclusion (15 CFR 734.7), for example for models beyond the best open model.
3. An order, or a training or deployment ban, after a crisis.

Two things could raise the odds:

• A Mythos-class Chinese open model (GLM-5.3) is already out (Anthropic).
• China's draft controls on top-10 models (The Decoder) could weaken the "we'd just cede ground to China" argument.

Model

I built quarterly hazards in Python and accumulated them as 1 − ∏(1 − h).

• Trump era: 0.2% for Q4 2026, then 0.3–0.5% a quarter, about 3% in total through 2028.
• 2029–31 mixture:
  • Democratic president (62%, shaded from Polymarket's 65%): about 1.05–1.35% a quarter, or 13% conditional.
  • Republican president: about 0.33% a quarter, or 4% conditional.
• Result: 1.7% by end-2027, 3.0% by end-2028, 6.5% by end-2029, 9.6% by end-2030 and 12.3% by end-2031.

Markets are broader and point low near term:

• Kalshi, "US bans any open-source model in 2026": bid 5¢, ask 9¢.
• Polymarket, same question: 7%. Both include bans on Chinese models, which this question excludes.
• Kalshi, Trump action "establishing or directing" mandatory pre-release review by 1 Jan 2027: bid 8¢, ask 15¢.

Consistency. This sits below the sister forecasts for pre-release access (45%) and a G20 government barring the public from capable models (15%), and near a joint approval body (11%).

Strongest case against

• Higher (about 20–25%). This world is moving fast: AGI claims, rogue-agent incidents, labs calling for slowdowns. A mass-casualty or major cyber incident traced to downloadable weights could quickly produce a law or rule, and any frontier licensing regime would naturally cover open release.
• Lower (about 5%). Every power centre, including Democrats, prefers testing, liability and action against Chinese models, and US limits cannot stop Chinese releases.

I stay near 12%.

Reading adopted: literal. It must be an explicit licence, approval or prohibition on publishing weights above a stated threshold, or a pre-release approval regime that covers open release. Company-specific letters, emergency suspension powers and implicit kill-switch effects do not count. This question has not been forecast before.

Question. By the last day of each quarter, has the US federal government required a licence or other government approval before the weights of AI models above a stated capability or compute threshold can be published openly?

Definitions.

• Required: imposed by an enacted federal statute, a final or interim final rule in the Federal Register (including amendments to the Export Administration Regulations, EAR), or a signed executive order that itself establishes the obligation, even with a later compliance date. A measure that only authorises or directs an agency to create such a requirement does not count.
• Licence or other government approval: a federal licence, authorisation, permit or clearance that the issuing body may refuse or condition, or an outright prohibition. An advance-notice requirement counts only if the government may block publication during the notice period.
• Published openly: made publicly downloadable, under any licence. Export controls count if they reach such publication, for example by taking covered weights out of the EAR's exclusion for published technology (15 CFR 734.7) or by treating publication as an export.
• Stated capability or compute threshold: covered models are defined by training compute, parameter count, evaluation or benchmark results, demonstrated abilities, or a formally designated tier defined by such criteria. A list of named models or developers alone does not count.

Resolves YES if a qualifying official publication establishes the requirement for at least some models above the threshold, applying to US developers or US persons generally. No denied application or enforcement is needed.

Resolves NO / Exclusions.

• Registration, notification without blocking power, reporting, safety frameworks, and government testing or access before release without a power to refuse. The Artificial Intelligence Risk Evaluation Act (S.2938) as introduced would not qualify: it bars deployment, including open-source release, without taking part in an evaluation program, but gives no power to refuse.
• Controls that exclude published weights, such as ECCN 4E091 in the January 2025 AI Diffusion rule; Executive Order 14409 (2 June 2026).
• Company-specific directives, "is informed" letters and unpublished orders.
• Restrictions only on importing, hosting or using foreign-developed models, or on particular content or uses; conditions of federal contracts, grants or compute; state laws, bills, proposed rules and guidance.

Resolving source. Congress.gov and GovInfo (statutes), the Federal Register (rules and orders) and the White House's official publication (orders), narrower than Public confirmation (A2). Apply the Resolution window and Resolution date (A3), dating YES by first official publication, not the compliance date. Later stays, vacatur or repeal do not reverse a YES.