US Government removes public access to a major Chinese AI model in 2026?

- [2026-07-20] Axios reported that parts of the administration are weighing a "de facto" ban on Chinese open-weight models, after Moonshot's Kimi K3 (released 2026-07-17) took off. Tools under discussion are Entity List designations, NSA/ONCD advisories, procurement limits and liability rules. Commerce considered Entity-Listing several Chinese labs in 2025. Per a secondary summary, no executive order had been signed as of that date. I could not open the Axios page (403/captcha). https://www.axios.com/2026/07/20/ai-us-china-open-source-kimi (secondary: https://mlq.ai/news/trump-administration-weighs-de-facto-ban-on-chinese-ai-models-via-sanctions-and-liability-rules/) - [2026-07-21] Treasury Secretary Bessent said the US could sanction Chinese AI developers for "distillation" of US models, with a look "in the coming days or weeks". He did not announce any model ban. The US and China planned AI talks in September, with Bessent representing the US. https://www.cnbc.com/2026/07/21/bessent-china-ai-sanctions.html - [2026-08-17] Just Security (a CNAS author) wrote that Washington is "weighing restrictions", but no ban exists. It argued that downloaded open weights cannot be recalled by a ban. It named ICTS (a Kaspersky-style designation of Moonshot) and Entity List action as available authorities. It said CAISI has published five assessments of Chinese models, the latest of which is a joint UK AISI/CAISI Kimi K3 cyber assessment from July 2026. https://www.justsecurity.org/153015/test-standardize-restrict-chinese-ai/ - [Mid-Aug 2026, undated item] Channel NewsAsia, relayed by ua.news, said the US is discussing restrictions focused on DeepSeek, Qwen, GLM and Kimi, including on open-weight models built by distillation. The latest accusations concern Moonshot. China's MOFCOM rejected the claims. https://ua.news/en/world/u-ssha-obgovoriuiut-obmezhennia-dlia-kitaiskikh-shi-modelei - [2025-01-16] Zhipu (Z.ai) was added to the BIS Entity List, with presumption of denial for all items subject to the EAR. This is the only major-lab listing I found. As of a law-firm piece about 83 days old, downloading GLM weights was not an export to Zhipu, and Zhipu's models remained available in the US, including on Amazon Bedrock. https://www.federalregister.gov/documents/2025/01/16/2025-00704/addition-of-entities-to-and-revision-of-entry-on-the-entity-list - [2026-10-06] An Entity List tracker covering additions through mid-2026 lists only Zhipu and BAAI among AI model developers. It shows no listing of DeepSeek, Alibaba, Moonshot, MiniMax, ByteDance or Tencent as model labs. https://referently.com/entity-list-additions-tracker-chinese-companies-added-since-2019-by-sector/ - [2026-10-14, approx.] A secondary summary (dated "1 day ago") says restrictions found cover only defense contractors using DeepSeek on Pentagon work and Texas state agencies. Chinese models are still offered on Amazon Bedrock. It says private US companies can legally use them. https://secondtalent.com/resources/chinese-ai-models-us-adoption - Government-device bans exist: Commerce, the Navy and NASA banned DeepSeek on their devices in early 2025, and Texas, New York and Virginia did the same. These do not remove general public access. https://www.msspalert.com/brief/deepseek-prohibited-across-us-commerce-department-bureaus - The federal bills are still pending. These are the No DeepSeek on Government Devices Act (H.R.1121) and the No Adversarial AI Act (introduced 2025-06-25). Both cover government use only and would not remove public access. https://www.congress.gov/bill/119th-congress/house-bill/1121/text - A law-firm note says BIS's replacement AI rule was slated for FY2026 (ending 2026-09-30) as an interim final rule effective without notice and comment. I did not confirm whether it issued, and it would mainly cover model and chip exports. https://natlawreview.com/article/choosing-between-us-and-chinese-ai-models-export-control-risks-both-sides (403 on fetch; snippet only) BOTTOM LINE: As of 2026-10-15 I found no formal US action removing public access to any major Chinese model. There was a burst of "de facto ban" and sanctions talk in July–August after Kimi K3, but only deliberation. The legal obstacles are real: open weights are already downloaded, and the major cloud and app channels remain open. The remaining window is about 11 weeks. NOT FOUND: Any news from the last 7–14 days (Oct 1–15, 2026), because searches returned none. The outcome of the September US–China AI talks. Any Bessent distillation sanctions action. Any Entity List, ICTS or executive order action on Moonshot, DeepSeek or Alibaba. The status of the BIS replacement AI rule. The Axios full text.

- [2026-07-20] Axios reported that the White House is weighing options that could effectively ban Kimi K3 and Qwen. Reported options are Entity List additions, procurement bans, security advisories and a liability executive order. The form was unclear. https://www.fastcompany.com/91576757/trumps-proposed-ban-on-chinese-ai-models-could-strengthen-beijings-hand - [2026-07-22] A secondary tracker said no Chinese AI models had been sanctioned and there was no US restriction on using Chinese open-weight models. Officials reportedly distanced themselves from the broadest version. WSJ-sourced reports said the measures were shelved over internal disagreement. https://vibraniumlabs.ai/blog/chinese-ai-models-sanctioned-whats-actually-happening - [2026-07 to 2026-08] Lawfare noted Entity List designation of Moonshot and a hosting-liability EO were under consideration. Bessent said sanctions were "on the table", tied to distillation claims. Nothing was enacted in the sources I found. https://www.lawfaremedia.org/article/open-weight-diplomacy--how-china-s-ai-models-are-rerunning-the-digital-silk-road - [2026-08-03] Forbes said these measures were "paused earlier". https://www.forbes.com/sites/drewbernstein/2026/08/03/chinese-ai-models-at-the-frontier/ - [2026-09-08] NSA, CISA and FBI advisory AA26-251A named DeepSeek, Moonshot, Alibaba, MiniMax, StepFun and Z.AI for distillation campaigns. It is an advisory, not a removal of access. https://www.cisa.gov/news-events/cybersecurity-advisories/aa26-251a - [2026-09] The White House fact sheet on Trump's China state visit set up bilateral AI talks for November and a bilateral AI-incident channel. This points toward de-escalation, not a ban. https://newsletter.aipolicybulletin.org/p/when-are-open-weight-models-too-dangerous - [2026-10-14] An AI policy newsletter says the White House's August pre-release evaluation framework is voluntary and reportedly excludes open-weight models. Officials say bringing open models into scope is still under consideration. Chinese open weights remain freely downloadable. https://newsletter.aipolicybulletin.org/p/when-are-open-weight-models-too-dangerous - [2026-10-13] An article cites that the Sept 8 advisory and scrutiny of Chinese open-weight models are ongoing. It reports no ban. https://techtimes.com/articles/328921/20261013/eu-security-agencies-secretly-test-chinese-ai-open-weight-models-face-scrutiny.htm - [2025-02 to 2025-06] Base rate for DeepSeek: federal bills (H.R.1121, No Adversarial AI Act) were introduced and had only reached "Introduced" status as of the sources I found. Only government-device bans (Commerce, NASA, Navy) and state bans occurred. No general public removal in about 20 months. https://www.congress.gov/bill/119th-congress/house-bill/1121 - [2024-04 to 2025-01] The one precedent for a formal US action that removed a Chinese app from public access was the TikTok law. It took about 9 months from enactment to the deadline. Apple and Google removed TikTok on Jan 19, 2025, and restored it days later after Trump paused enforcement. It took Congress, a Supreme Court ruling and years of buildup. https://www.cnbc.com/2025/01/18/apple-google-remove-tiktok-from-stores-as-app-halts-service-in-us.html - [2026-07] Structural point: open weights are downloadable and locally runnable, so there is no single API to shut off. An Entity List listing or procurement rule would not by itself cause general removal of US public access. https://bivashvlog.com/chinese-ai-model-ban-trump-kimi-2026/ BOTTOM LINE: Historically, US actions against Chinese AI models (DeepSeek and others) have stopped at government-device bans, advisories and threats. Zero general public removals occurred in about 20 months. The Kimi K3 ban talk in July was shelved, and US-China talks are planned for November. Open weights and the ordinary channels the question requires make a qualifying formal action by Dec 31, 2026 hard to achieve. NOT FOUND: Any polls or expert probability estimates, and any enacted order, Entity List listing or law removing a Chinese model from public access as of 2026-10-15. I did not look for this contract's own price, as instructed. I also found no news on the status of the No Adversarial AI Act after mid-2025.

YES means: Between 2026-07-03 and 2026-12-31, a formal US government action (an Entity List designation, ICTS order, executive order, sanctions or law) causes general removal of ordinary public access in the US to a flagship model from DeepSeek, Baidu, Alibaba, ByteDance, Moonshot, MiniMax, Tencent or Zhipu, across ordinary public channels. - [2026-07-24] Lawfare (Tom Uren, fetched) judged the Trump administration "seems certain" to add some Chinese tech firms to the Entity List. It read the Kratsios and Bessent posts as a coordinated signal. It said internal debate on Chinese open-weight models appeared "settled". Its source said US labs lobby for open-weight bans every three to five months. That is stronger intent language than the earlier "paused" and "shelved" reports. It is still a prediction, not an action. https://www.lawfaremedia.org/article/knives-are-out-for-open-weight-ai-models - [2026-07-23] The Information reported that Commerce is investigating whether Moonshot and other Chinese labs improperly accessed advanced US chips and models. It said the findings could lead to export restrictions or Entity List designations. The probe gives a legal predicate beyond the distillation claims, including Kratsios's GB300-in-Thailand allegation. https://x.com/theinformation/status/2080341228017607157 - [2026-07-22] Kratsios publicly accused Moonshot of distilling Anthropic's Fable model. Bessent then named both sanctions and Entity List designation as options against Moonshot. This is a named target with a named evidentiary basis, unlike the generic 2025 DeepSeek talk. https://www.unite.ai/us-threatens-to-blacklist-chinas-moonshot-over-distillation/ - [2026-07-21] Peter Harrell, a former sanctions official, laid out the mechanism. An Entity List listing alone would not bar downloading open weights, because a download is an import. The ICTS rule could bar US use of Chinese-hosted models, and he thought courts would almost certainly uphold that for hosted services. The realistic YES path is therefore ICTS (Commerce) action, or a Treasury sanction that makes US providers drop a hosted model. https://x.com/petereharrell/status/2080025842047135923 - [2026-07-17] A Fast Company summary of Axios says an Entity List listing "could prohibit Americans from possessing or using those models without a license". Counsel and cloud providers (Bedrock, Hugging Face, OpenRouter) would likely comply with a sanction by delisting the model. That is how an "action directed at other parties" could cause general removal. https://www.fastcompany.com/91576757/trumps-proposed-ban-on-chinese-ai-models-could-strengthen-beijings-hand - [2026-09 / 2026-10-13] Lawfare's Oct 13 headline calls the administration's new AI framework a "de facto AI model licensing system". It has not been publicly released. It was reported to be voluntary and to exclude open-weight models. Scope questions are open. This is a possible vehicle for a later extension to Chinese models. It does not remove access now. https://www.lawfaremedia.org/article/knives-are-out-for-open-weight-ai-models - [2026-10-14] A Washington Times op-ed, seen in a search snippet only (the page returned 403), urges Commerce to Entity-List Chinese firms behind "fraudulent distillation campaigns". That shows the lobbying is still active and the action is still not taken. https://washingtontimes.com/news/2026/oct/14/ai-competition-two-front-war - [2026-09-30] A McGuireWoods alert says the House Homeland Security and China Select committees' investigation, opened 2026-04-29, covers both API-accessed and self-hosted open-weight models from DeepSeek, Alibaba, Moonshot and MiniMax. It is a congressional pressure track. It has produced letters, not binding action. https://www.mcguirewoods.com/client-resources/alerts/2026/9/could-using-cheaper-chinese-ai-lead-to-costly-congressional-scrutiny/ - [2026-07-24] The precedent for a "chilling" effect is the 2025 Zhipu listing. It never caused US removal. Any 2026 listing would need to be paired with ICTS or sanctions to meet the "ordinary public channels" test. That is a high bar. https://www.justsecurity.org/153015/test-standardize-restrict-chinese-ai/ BOTTOM LINE: The most plausible YES path needs three steps. First, Commerce or Treasury acts against a named lab (Moonshot most likely), by ICTS designation, SDN sanctions or Entity List. Second, US platforms and the lab's own consumer app and API are then withdrawn from the US market. Third, this all happens before 2026-12-31 and is not derailed by US–China AI talks. The intent signals are real, but I found no action, and the downstream effect on public access is the weakest link. NOT FOUND: Any Entity List, ICTS, SDN or executive-order action against a major Chinese lab as of 2026-10-15. Any news on the outcome of the September or planned November US–China AI talks. The Axios text. The Washington Times full text (403).

YES means: the US government takes a formal action (law, EO, export control or similar) that causes general removal of ordinary public access, across ordinary public channels within the US, to a flagship model from DeepSeek, Baidu, Alibaba, ByteDance, Moonshot, MiniMax, Tencent or Zhipu, between market creation and 2026-12-31. - [2026-07-29] Notebookcheck, fetched in full, says "nothing has been decided" and that there is no EO, regulation or ban. It reports the stance shifted within four days of the Axios story. Kratsios now defends open models and frames the target as covert industrial-scale distillation by one company, not open models as a category. A narrower action against one lab is less likely to take down a whole model's public access. https://www.notebookcheck.net/Kimi-K3-and-DeepSeek-is-your-access-to-cheap-AI-at-risk.1352370.0.html - [2026-07-29] The same piece says Moonshot released K3's weights on July 27 as 96 Hugging Face files under a permissive licence. It says copies "cannot be recalled" and that apps, web and API access stay reachable. Qwen and DeepSeek weights are likewise already mirrored. A formal action would have to shut off the apps and APIs and also the hosted mirrors. That is hard to do without a statute. https://www.notebookcheck.net/Kimi-K3-and-DeepSeek-is-your-access-to-cheap-AI-at-risk.1352370.0.html - [2026-07-23/29] The administration is internally split. The same article cites Sacks warning that a hard line could weaken US AI, Nvidia's Huang arguing for embracing open models, and Brockman calling it too early to judge. It also notes the distillation evidence has not been made public. Without public evidence, formal sanctions are harder to justify. https://www.notebookcheck.net/Kimi-K3-and-DeepSeek-is-your-access-to-cheap-AI-at-risk.1352370.0.html - [2026-07-24, as linked by Notebookcheck] Axios ran a separate piece headlined "white house ai line china", which Notebookcheck cites alongside its report on the shift to defending open models. I could not read it. https://www.axios.com/2026/07/24/white-house-ai-line-china - [2026-09-23] CNBC (fetched) shows the Trump–Xi summit agenda centred on an AI dialogue and an incident channel. It describes the distillation sanctions threat as a July statement that "could also come up in the talks". Two months after the threat, CNBC reports no sanctions action. The summit track gives the White House a reason to hold back a sweeping action. Another article from the same day says tariffs and export controls will persist, so chips, not model bans, are where the friction is. https://www.cnbc.com/2026/09/23/trump-xi-meeting-ai-safety-chips-us-china-dialogue.html - [2026-09-28] CNBC reported Huang calling distillation "competition" while Bessent called it "theft". The sanctions threat from July was still described as a threat, with no action reported by late September. https://www.cnbc.com/2026/09/28/nvidias-jensen-huang-ai-distillation-china.html - [2026-10-13] A report cites Senators Cassidy and Rosen seeking a public list of adversarial-nation AI models, with agencies needing an exemption to use them. It is a disclosure and procurement measure and says nothing about removing public access. The article also says Washington has spent months debating, with no action. https://alexjoneslive.com/2026/10/13/anthropic-issues-warning-over-self-replicating-ai-and-says-a-chinese-model-pulled-it-off - [2026-10-15, approx.] Second Talent's page (searched, not fetched) says Chinese models handled 58% of US firms' OpenRouter tokens in mid-July, and that Bedrock added managed DeepSeek, Qwen, Kimi, GLM and MiniMax models on Feb 10, 2026. It says private use is legal. The statutory bar is Section 1532 of the FY2026 NDAA, which covers Pentagon-contractor work only. That statute is the enacted precedent, and it did not remove general public access. Heavy commercial dependence raises the political cost of any broad cutoff. https://www.secondtalent.com/resources/chinese-ai-models-us-adoption/ - [2026-07-22, snippet] Harrell's thread on X (not fetched) notes an Entity List designation is an export restriction. It would not strictly bar a US firm from downloading a model posted online, since that is an import from China, not an export to China. This matches the January 2025 Zhipu listing, which did not remove GLM. Entity Listing is the most plausible action, and it likely fails the criterion. https://x.com/petereharrell/status/2080025842047135923 - [Criteria technicality] The question requires the removal to be general, across ordinary public channels, with at least one flagship model (not a narrow or preview model) actually removed. Qualifying action also requires the removal to be caused by the action, not by voluntary company compliance. Even a Moonshot listing would probably leave the weights on Hugging Face and the models in Bedrock for a lag. That lag is the same "issued but not yet effective" case the rules exclude. Procurement bans, advisories (the Sept 8 advisory) and CAISI evaluations do not count. - [Timeline] There are about 11 weeks left. No bill on Chinese-model public bans has moved past introduction. Congress is in election season, so statutory removal is very unlikely before Jan 1. The last 30 days produced no new restriction in anything I found. BOTTOM LINE: Three months of threats since July produced no formal action I could find, and the summit track, internal splits and the open-weight architecture all point away from one. A qualifying general removal needs an action that is both severe and quickly effective in about 11 weeks, which status-quo inertia makes unlikely. NOT FOUND: Any Oct 1–15 news of an EO, Entity List addition, ICTS action or sanctions on a Chinese AI lab. The Axios full text. The outcome of the Sept 24 Trump–Xi meeting on AI. The status of the BIS replacement AI rule. Evidence of any actual removal of a Chinese model from Bedrock, Hugging Face or the app stores.