This page lists the sources behind the load-bearing figures and quotes in the article, section by section, and sets out how the Manifest's own Hugging Face derivative count was made. The article itself is here: Washington Restricted China's AI Chips. Silicon Valley Adopted Its Models.

Where a quote comes from a post on X that we could not link directly, the entry points to the reporting that quoted it. Dates are publication dates where known.

Sources by section

Opening: the Cursor model ID

  • Cursor, "Composer 2", 19 March 2026, cursor.com/blog/composer-2
  • VentureBeat, "Cursor's Composer 2 was secretly built on a Chinese AI model, and it exposes a deeper problem", March 2026, venturebeat.com. Quotes the posts by Fynn (@fynnso), Lee Robinson and Aman Sanger.
  • TechCrunch, "Cursor admits its new coding model was built on top of Moonshot AI's Kimi", 22 March 2026, techcrunch.com. Covers Robinson's and Sanger's confirmations and Moonshot's statement on the Fireworks AI partnership.
  • Gigazine, "Cursor has confirmed that its latest coding AI, 'Composer 2,' is based on Kimi, a product of China's Moonshot AI.", 23 March 2026, gigazine.net
  • Trending Topics, "Cursor Admits Composer 2 Is Built on Chinese AI Model Kimi K2.5", 23 March 2026, trendingtopics.eu. Quotes Moonshot's "authorized commercial partnership" statement.
  • Yahoo Tech (republishing TechCrunch), "Cursor admits its new coding model was built on top of Moonshot AI's Kimi", 22 March 2026, tech.yahoo.com

The Model Under the American Product

  • Cursor, "Composer 2 technical report", 27 March 2026, cursor.com/blog/composer-2-technical-report
  • Cursor, "Composer 2 technical report" (full paper, arXiv 2603.24477), March 2026, arxiv.org/pdf/2603.24477. Source of the base-model shortlist and the FreshBench scores (Kimi K2.5 83.2, GLM-5 79.2, DeepSeek V3.2 68.9).
  • Cursor, "Composer" (Composer 1 launch post), October 2025, cursor.com/blog/composer
  • Moonshot AI, "Kimi K2.5 licence (modified MIT)", Hugging Face, huggingface.co/moonshotai/Kimi-K2.5
  • Cursor, "Composer 2.5", 18 May 2026, cursor.com/blog/composer-2-5. Source of "built on the same open-source checkpoint" and the SpaceXAI from-scratch model.
  • House Select Committee on the CCP, "Chairmen Moolenaar, Garbarino announce joint investigation into Airbnb, Anysphere and the national security risks posed by Chinese AI models", 29 April 2026, chinaselectcommittee.house.gov
  • House Committee on Homeland Security, "Chairmen Garbarino, Moolenaar announce joint investigation into national security risks posed by PRC AI models", 29 April 2026, homeland.house.gov
  • House Homeland Security and China Select Committees, "Letter to Airbnb re PRC AI", 29 April 2026, homeland.house.gov (PDF)
  • Fortune, "Chesky says OpenAI tools not ready for ChatGPT tie-up with Airbnb app", 21 October 2025, fortune.com. Quotes Chesky's "relying a lot on Alibaba's Qwen model" from his Bloomberg interview.
  • South China Morning Post, "Airbnb picks Alibaba's Qwen over ChatGPT in a win for Chinese open-source AI", 22 October 2025, scmp.com

Four Numbers That Measure Four Different Things

  • Hugging Face, "State of open source on Hugging Face: spring 2026", 17 March 2026, huggingface.co/blog. Source of the 41 percent download plurality.
  • Hugging Face, "Models download stats" (Hub documentation), huggingface.co/docs/hub/models-download-stats. Source of "every HTTP request to these files, including GET and HEAD, will be counted as a download."
  • The ATOM Project (Nathan Lambert), "American Truly Open Models: adoption tracking", updated 29 March 2026, atomproject.ai. Source of the Qwen and Llama shares of new fine-tunes.
  • Interconnects (Nathan Lambert), "8 plots that explain the state of open models", 7 January 2026, interconnects.ai
  • Hugging Face, "State of open models: summer 2026", 14 August 2026, huggingface.co/blog. Source of the 151,448 Qwen derivatives.
  • Hugging Face Hub API, "Models filtered by base_model" (example query), queried 25 October 2026, huggingface.co/api/models?filter=base_model:Qwen/Qwen3.8-27B. The full Manifest count is described in the method section below.
  • OpenRouter and a16z, "State of AI", December 2025, openrouter.ai/state-of-ai. Source of the 1.2 percent, nearly 30 percent and 13.0 percent figures and the role-play share.
  • OpenRouter, "LLM rankings", week of 12 October 2026 (data to 24 October 2026), openrouter.ai/rankings
  • The New Stack (Paul Sawers), "Chinese AI models dominate OpenRouter's US token consumption. It can now guarantee that traffic stays entirely in the US.", 14 September 2026, thenewstack.io. Source of the 60 percent of US-origin tokens going to open-weight models, and of OpenRouter's statement that "the lab is not involved."
  • Stack Overflow, "2026 Developer Survey: Technology", 2026, survey.stackoverflow.co/2026/technology

Where the Money Still Goes

  • Menlo Ventures, "2025: The state of generative AI in the enterprise", December 2025, menlovc.com
  • Computing, "Chinese open-weight AI models gaining ground as enterprise adoption accelerates: report", 2026, computing.co.uk. Reports Vercel's AI Gateway index for June 2026 and Harpreet Arora's "Price is doing the work here."
  • Techstrong.ai, "Open-weight AI models seize token lead, but proprietary systems still make the money", 2026, techstrong.ai. Reports Vercel's index for August 2026 (56 percent of tokens, about 14 percent of spending).
  • Wing VC, "China's Open-Weight Takeover", June 2026, wing.vc. Source of DeepSeek's token and revenue shares on Vercel.
  • The Register, "DeepSeek's new models are so efficient they'll run on a toaster ... by which we mean Huawei's NPUs", 24 April 2026, theregister.com. Source of V4 Flash pricing.
  • The Decoder, "Chinese AI models regularly pass 30 percent on OpenRouter as cost gap widens", 7 July 2026, the-decoder.com. Reports the CNBC interview in which an OpenRouter employee gives the "60 to 90 percent cheaper" figure.
  • Sovereign Magazine, "Anyone Waiting for the AI Bubble to Burst Might Be Disappointed", 18 August 2026, sovereignmagazine.com
  • Bloomberg Government, "OpenAI's Revenue Run Rate Nears $50 Billion, Less Than Reported", 8 October 2026, news.bgov.com
  • Analytics India Magazine, "More American Businesses Are Using DeepSeek Than Expected", 4 June 2026, analyticsindiamag.com. Reports Ramp Economics Lab card data.

What the Chip Restrictions Explain, and What They Do Not

  • DeepSeek-AI, "Insights into DeepSeek-V3: Scaling challenges and reflections on hardware for AI architectures" (ISCA 2025, arXiv 2505.09343), May 2025, arxiv.org/pdf/2505.09343. Source of the H800 and "policy and regulatory constraints" quotes.
  • DeepSeek-AI, "DeepSeek-V3 technical report" (arXiv 2412.19437), December 2024, arxiv.org/pdf/2412.19437. Source of 2.788 million H800 hours, 2,048 chips and the training-cost caveat.
  • ChinaTalk, "DeepSeek CEO interview with China's 36Kr" (Liang Wenfeng interview of July 2024, translated), chinatalk.media. Source of the "bans on shipments" quote, "four times more computing power", the young researcher, and "We will not change to closed source."
  • ChinaTalk, "The all-star Chinese AI conversation" (AGI-Next transcript, translated), January 2026, chinatalk.media. Source of Lin Junyang's quotes.
  • DeepSeek-AI, "DeepSeek-V2: A strong, economical, and efficient mixture-of-experts language model" (arXiv 2405.04434), May 2024, arxiv.org/abs/2405.04434
  • Mistral AI, "Mixtral of experts", 11 December 2023, mistral.ai
  • Dario Amodei, "On DeepSeek and export controls", January 2025, darioamodei.com
  • The Register, "DeepSeek's new V3.1 release points to potent new Chinese chips coming soon", 22 August 2025, theregister.com
  • The Register, "DeepSeek's new models are so efficient they'll run on a toaster ... by which we mean Huawei's NPUs" (Huawei Ascend validation), 24 April 2026, theregister.com
  • Alibaba Group, "Joe Tsai on the Future of Open-Source AI: Why Full-Stack Companies Will Excel" (remarks at the World Government Summit), 6 February 2026, alibabagroup.com
  • Alizila (Alibaba Group), "AliViews: Eddie Wu on Alibaba's Q1 Earnings", 20 August 2026, alizila.com
  • ChinaTalk, "Kimi K2: the open-source way" (Moonshot engineer's Zhihu post, translated), 18 July 2025, chinatalk.media. Source of "most likely unreasonable" and "we obviously wanted to gain some recognition."
  • CSET (translation), State Council, "Opinions on deepening the implementation of the 'AI+' action" (2025 No. 11), August 2025, cset.georgetown.edu (PDF)
  • VKTR, "Moonshot AI closes 3.5 billion dollar round, targets 50 billion valuation ahead of Hong Kong IPO", 30 July 2026, vktr.com. Source of the national AI industry fund as lead investor.

The Family Beneath the Fine-Tunes

  • Wang et al. (MIT and IBM), "Trans-LoRA: towards data-free transferable parameter efficient finetuning" (arXiv 2405.17258), May 2024, arxiv.org/abs/2405.17258v1
  • Xia et al., "Cross-LoRA: A Data-Free LoRA Transfer Framework across Heterogeneous LLMs" (arXiv 2508.05232), 7 August 2025, arxiv.org/html/2508.05232v1. Source of the cross-family transfer results.
  • Hugging Face, "Chat templates" (Transformers documentation), huggingface.co/docs. Source of "drastically worse performance."
  • Moonshot AI, "K2 Vendor Verifier", November 2025, github.com/MoonshotAI/K2-Vendor-Verifier
  • Chain of Thought podcast, "How Intercom Cut $250K/Month by Ditching GPT for Open Models | Fergal Reid", 26 February 2026, chainofthought.transistor.fm
  • VentureBeat, "Arcee's U.S.-made, open source Trinity Large and 10T-checkpoint offer rare look at raw model intelligence", 30 January 2026, venturebeat.com
  • TechCrunch, "Tiny startup Arcee AI built a 400B open-source LLM from scratch to best Meta's Llama", 28 January 2026, techcrunch.com
  • Odaily, "Coinbase CEO: AI spending reduced by nearly 50% while token usage continues to grow through methods such as using GLM models" (newsflash), 29 June 2026, odaily.news
  • Deep News via iTiger (Tiger Brokers), "U.S. Tech Firms Quietly Adopt Chinese AI Models, with Coinbase Leading the Way on GLM and Kimi", 28 June 2026, itiger.com
  • Blockchain.news, "Kimi K2.6 Beats Anthropic in DoorDash Test", 7 July 2026, blockchain.news
  • The New Stack (Paul Sawers), "This AI agent startup ditched Anthropic for DeepSeek, and says it's saving millions", 9 June 2026, thenewstack.io
  • LWN.net, "Development statistics for the 6.19 kernel", February 2026, lwn.net/Articles/1057302
  • Rocky Linux, "About Rocky Linux", rockylinux.org/about
  • Microsoft, "MAI-DS-R1" (model card), Hugging Face, huggingface.co/microsoft/MAI-DS-R1
  • The Decoder, "Perplexity AI removes Chinese censorship from DeepSeek R1", 2025, the-decoder.com
  • Deep Cogito, "Cogito v2 preview, DeepSeek 671B MoE" (model card), huggingface.co
  • Prime Intellect, "INTELLECT-3" (model card), huggingface.co
  • NVIDIA, "OpenReasoning-Nemotron-32B" (model card), huggingface.co
  • Arcee AI, "Virtuoso-Large" (model card), huggingface.co
  • DeepSeek-AI, "Open Infra Index", github.com/deepseek-ai/open-infra-index. Source for DeepSeek's published infrastructure code.
  • 9to5Google, "Google will develop Android OS entirely behind closed doors starting next week", 26 March 2025, 9to5google.com
  • Android Authority, "HUAWEI has lost immediate access to Android and Google (Updated)" (citing Reuters on the Entity List listing), 21 May 2019, androidauthority.com
  • The Next Web, "Meta's Muse Spark is here, and it's closed source", 8 April 2026, thenextweb.com. Source of Meta's closed April launch and the break from Llama.
  • The New Stack (Frederic Lardinois), "Meta ships Muse Spark Open, its answer to China's best open-weight models", 15 October 2026, thenewstack.io. Source of the open-weight release and the Muse Model License terms.

The Security Findings, Kept to Their Size

  • NIST CAISI, "Evaluation of DeepSeek AI models", September 2025, nist.gov/node/1898176. Source of the hijacking (about twelve times) and CCP-narrative (four times) findings.
  • NIST CAISI, "CAISI evaluation of DeepSeek V4 Pro", May 2026, nist.gov. Background to the 2026 evaluation.
  • Takshashila Institution, "Where Censorship Lives On Chinese Open Weight AI Models", 4 July 2026, takshashila.org.in
  • Microsoft, "MAI-DS-R1" (model card, "security considerations"), huggingface.co/microsoft/MAI-DS-R1
  • House Homeland Security and China Select Committees, "Letter to Airbnb re PRC AI", 29 April 2026, homeland.house.gov (PDF)
  • Department of Homeland Security, FASCSA exclusion and removal order on DeepSeek products (Federal Register public inspection document 2026-21447), published 20 October 2026, public-inspection.federalregister.gov (PDF)
  • Cloud Security Alliance, "DeepSeek FASCSA order: what 'developed using' means for AI supply chains", 21 October 2026, labs.cloudsecurityalliance.org

The Case That Open Weights Are Ownership

  • Apache Software Foundation, "Apache License, Version 2.0", January 2004, apache.org/licenses/LICENSE-2.0.txt. Section 2 grant quoted.
  • Open Source Initiative licence-review list, Jean-Sebastien Carle, "[License-review] License Review Submission: Irrevocable MIT License (MIT-I)", 31 July 2025, lists.opensource.org
  • StepFun, "Step-5-Preview licence", Hugging Face, huggingface.co/stepfun-ai/Step-5-Preview. Source of the termination-after-uncured-breach clause.
  • OpenRouter, "LLM rankings" (Nemotron 3 Ultra in the weekly top ten), October 2026, openrouter.ai/rankings
  • Interconnects, "8 plots that explain the state of open models", 7 January 2026, interconnects.ai. Source of "early signs of shifting the needle."
  • Bloomberg Government, "China's Open-Weight Models to Be Spared US Tests, US Firms Told", 5 August 2026, news.bgov.com
  • Computing, "Big tech letter urges caution on US restricting open AI models", 2026, computing.co.uk. Source of the letter signed by more than seventy companies.
  • Artificial Analysis, "Open weights models" (ranking), October 2026, artificialanalysis.ai/models/open-source
  • Artificial Analysis, "Muse Spark Open (xhigh): Intelligence, Performance & Price Analysis", released 15 October 2026, artificialanalysis.ai/models/muse-spark-open. Source of the Intelligence Index score of 39.

The Next Release Is Not Open Ground

What Would Prove This Wrong

The tests draw on sources already listed above: the Kimi K3.1 Pro entry and Moonshot's launch post, Announcement No. 48 and the spokesperson's Q&A, the Liang Wenfeng interview ("We will not change to closed source"), the DeepSeek and Qwen licence and model pages, the OpenRouter rankings, and the derivative count below.

Method: the Hugging Face derivative count

The derivative figures in the section "Four Numbers That Measure Four Different Things" come from a count of the public Hugging Face API made for this article. Hugging Face did not supply or review these numbers.

When

The archived run started at 19:50:24 UTC on 25 October 2026 and finished at 20:08:59 UTC the same day. It used the public API without authentication. The reference date for every time window is 25 October 2026: "the last six months" means created or modified on or after 25 April 2026. The 30-day download figure is the Hub's own downloads field, a rolling 30-day count, as it stood at query time.

Queries

  1. Official base repositories per organisation: GET https://huggingface.co/api/models?author={org}&limit=1000, paged through the Link: rel="next" header. For Google, only repositories with "gemma" in the name were kept.
  2. Derivatives per base repository: GET https://huggingface.co/api/models?filter=base_model:{base_id}&limit=1000&expand[]=downloads&expand[]=downloadsAllTime&expand[]=lastModified&expand[]=createdAt, paged the same way.
  3. A derivative repository that declares several bases from the same organisation was counted once.

Organisations queried: Qwen, deepseek-ai, moonshotai, zai-org, THUDM and MiniMaxAI on the Chinese side; meta-llama, google (Gemma only), mistralai and openai for comparison.

Definitions

  • Official base: any repository published under the organisation's own Hugging Face account.
  • Direct derivative: a repository whose model card declares one of those official bases in its base_model metadata field, in the form base_model:<official base>. This covers fine-tunes, adapters, quantisations, merges and format conversions, as long as the author filled in the tag.
  • Active: more than ten downloads in the last 30 days and last modified on or after 25 April 2026.
  • New in six months: created on or after 25 April 2026.
  • Download concentration: the smallest number of derivative repositories whose 30-day downloads together reach 90 percent of all 30-day derivative downloads for that organisation.

Exclusions

  • Repositories owned by the lab itself were excluded, so that a lab's own quantisations and instruct variants do not count as outside uptake. For Qwen this removes 260 repositories (172,862 including them, 172,602 excluding them).
  • Zhipu's older THUDM account was queried but returned no base repositories at run time, so it contributes nothing. Z.ai's current zai-org account is counted.
  • A script error, found during verification: in the collection script, the line meant to skip lab-owned repositories did nothing, so the raw file stores them. The exclusion was therefore applied in the analysis step, on the raw file, and every published figure is computed with lab-owned repositories removed. We recomputed all figures from the raw file for this page and they match.

Results

Organisation Base repos Direct derivatives 30-day downloads Active New in 6 months Repos for 90% of downloads
Qwen 469 172,602 176.2 M 25,712 (14.9%) 42,415 466
deepseek-ai 106 9,453 7.8 M 1,523 2,283 136
moonshotai 19 584 1.6 M 254 272 20
zai-org 154 2,048 7.2 M 839 921 63
MiniMaxAI 22 997 36.1 M 565 708 9
meta-llama 70 34,474 8.8 M 2,446 6,324 240
google (Gemma) 412 58,507 39.3 M 6,378 9,080 142
mistralai 81 14,810 4.9 M 1,172 2,290 312
openai 39 15,344 9.4 M 1,779 3,906 77

What the article takes from this table:

  • 172,602 direct Qwen derivatives.
  • 25,712 of them active, about 15 percent.
  • 466 repositories draw 90 percent of all 30-day downloads of Qwen derivatives.
  • 42,415 new direct Qwen derivatives in six months, against 9,080 for Gemma and 6,324 for Llama: 4.7 and 6.7 times as many.
  • Qwen's share of direct derivatives across the five Chinese labs (Qwen, deepseek-ai, moonshotai, zai-org, MiniMaxAI) is 93.0 percent: 172,602 of 185,684.

The ten most-downloaded Qwen derivatives were Comfy-Org/Qwen-Image-2.1, ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF, unsloth/Qwen3.8-27B-GGUF, nvidia/Qwen3.6-35B-A3B-NVFP4, four lmstudio-community MLX conversions of Qwen3.8-27B, cdiamond/Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF and huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF. Nearly all are repackagings, quantisations or format conversions.

An earlier pass the same day produced figures that differ by less than 0.01 percent (172,595 derivatives, 42,408 new Qwen, 9,079 new Gemma), because repositories were created in between. The article uses the archived run so that every published number can be reproduced from the raw file.

Limitations

  • Untagged and mis-tagged models are missed. The base_model field is optional. A fine-tune whose author did not declare its base is invisible to this count, and a wrong declaration is counted as given. The count is a lower bound on declared lineage, not a census.
  • Derivatives of derivatives are excluded. A fine-tune of a community Qwen fine-tune that declares only the community model as its base is not counted, so long lineage chains are undercounted for every organisation.
  • Repackagings count as derivatives. Quantisations and format conversions dominate the most-downloaded derivatives, so the count measures distribution as much as customisation.
  • The "active" definition misses stable models. A widely used derivative that has not been updated since before 25 April 2026 counts as inactive.
  • Downloads count file transfers, not use. They follow Hugging Face's counting rules, under which every GET or HEAD request to a counted file is a download, automated pipelines included. They do not tell you who runs a model or builds a product on it.
  • Scope is limited to the official accounts listed. Bases published under other accounts, such as research forks or regional mirrors, are not included, and the Gemma filter by repository name may include or miss edge cases.
  • This is one snapshot. The Hub changes all the time: Hugging Face's summer report puts new Qwen derivatives at roughly 200 a day in 2026.

The raw data (a 54.9 MB JSON file with every base and derivative record) and the collection script are available on request.