A short history of Qwen

If you're new here: strata→signal is a small workshop (plus a friendly dog with a white patch) that builds on its own machines and writes up what it measures. This page reads rather than measures; a bracketed number names a sentence's source, and every source but one was read on 2026-10-14 (UTC), the counters, licence files and cards again on 2026-10-21 and 2026-10-22; the one, Alibaba's press release on Qwen 4 [42], was read on 2026-10-22. Where we found no source, the claim is not here. The workshop has no business relationship with Alibaba or the Qwen team, which did not see this page before it went up.

The words this page leans on

• Open weights, and the API — model files anyone can download and run, against models that run only on the maker's servers, paid for by use.
• Parameters — the learned numbers inside a model; "27B" means about 27 billion. A mixture of experts uses only some of them for each token (a word or a piece of one), so "A3B" in a name means about 3 billion active.
• Apache 2.0 — a standard open-source licence; the copy shipped with Qwen3-8B grants a "no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute" the work [41].
• Derivative — a model made from another: a fine-tune, an adapter (a small add-on file), a quantization (a compressed copy) or a merge.

The family under many of our numbers

Several models this workshop has measured are Qwen models or are built on one. Of the 37 open models The cost of intelligence, as a utility read, the first, by the date its weights were published (2026-02-24), that fits a 24 GB card and clears its test against OpenAI's June 2023 flagship is qwen3.5:27b (a model not read could have crossed earlier); OpenJev, on Reading the answer, is "fine-tuned from Qwen3.8-27B". What Qwen models did on our cards lists each run; this page is the history behind them, not a review.

Who are these people, and where did they come from?

Qwen is short for Tongyi Qianwen (通义千问). The first Qwen model card calls Qwen-7B part of "the large language model series, Qwen (abbr. Tongyi Qianwen), proposed by Alibaba Cloud" [2]. The first technical report, posted on 2023-09-28 by 48 authors under the byline "Qwen Team, Alibaba Group", says the name "derives from the Chinese phrase Qianwen, which translates to 'thousands of prompts'" [6]. Alibaba Cloud is "the AI and cloud computing unit of Alibaba Group Holding", in the words of the South China Morning Post, which Alibaba owns [26].

Both Junyang Lin and Jingren Zhou are on that first author list [6]. CNBC called Lin "technical lead of Alibaba Cloud's Qwen team" on 2026-02-17 [22]. China Daily reported on 2026-03-05 that Lin had announced his departure on social media, writing "me stepping down. bye my beloved qwen", and that Yu Bowen, "head of post-training for the Qwen model", had also left. By the same report, an internal email from Alibaba's chief executive, Eddie Wu, said that Zhou, Alibaba Cloud's chief technology officer, "will continue to lead the Tongyi Lab", and that Alibaba "will continue to adhere to its open-source model strategy" [24].

The names on the licence files have changed too. The 2023 and 2024 custom licences define "We" as "Alibaba Cloud" [4, 11]; the Apache file shipped with Qwen3.8-27B reads "Copyright 2026 Alibaba Cloud" [32]; the Qwen Community License of August 2026 reads "Copyright (c) 2026 Qwen" [34]; the Qwen Research License dated 2026-09-20 defines "We" as "Hangzhou Tongyi Laboratory Technology Co., Ltd." [36]. None of the files says how those names relate.

How it began, in 2023

In April 2023, the Post reported, Alibaba Cloud "started a beta test of Tongyi Qianwen" [1]. On 2023-08-03 it released Qwen-7B and Qwen-7B-Chat on ModelScope and Hugging Face, two model-hosting sites [3], under the Tongyi Qianwen License Agreement, dated that day [4].

The licence let anyone use, copy and modify the weights, with two limits: a commercial user whose "product or service has more than 100 million monthly active users" must "request a license from Us", and "You can not use the Materials or any output therefrom to improve any other large language model (excluding Tongyi Qianwen or derivative works thereof)". It is governed by "the laws of China", with disputes before "The People's Courts in Hangzhou City" [4].

The technical report says the team "adopted the recent open-source approach of training large language models, LLaMA", Meta's model, with modifications, and trained the base models on "up to 3 trillion tokens" [6]. On 2023-09-13 the Post reported that Alibaba had opened Tongyi Qianwen to the public, after government approvals described below [1].

Qwen-14B followed on 2023-09-25, and Qwen-72B and Qwen-1.8B on 2023-11-30 [3]. The 1.8B shipped under a second text, the Tongyi Qianwen Research License, which grants its rights "FOR NON-COMMERCIAL PURPOSES ONLY" and defines non-commercial as "for research or evaluation purposes only" [5]. That November the team also posted Qwen-Audio, a model that takes sound as input [7].

The turn to Apache 2.0

Qwen1.5, on 2024-02-04, opened weights from 0.5 to 72 billion parameters; a 32B and a 110B followed in April 2024 [8, 37]. The three smallest went out under the research licence and the rest under the 2023 agreement [4, 5, 37]. Qwen2, on 2024-06-07, changed course: "While Qwen2-72B as well as its instruction-tuned models still uses the original Qianwen License, all other models, including Qwen2-0.5B, Qwen2-1.5B, Qwen2-7B, and Qwen2-57B-A14B, turn to adopt Apache 2.0!" [9]. Qwen2.5, on 2024-09-19: "All our open-source models, except for the 3B and 72B variants, are licensed under Apache 2.0" [10]. Its 72B carried a new Qwen License Agreement that kept the 100-million-user clause and replaced the ban on improving other models with a duty: whoever uses the model or its outputs "to create, train, fine-tune, or improve an AI model that is distributed or made available" must "prominently display 'Built with Qwen' or 'Improved using Qwen'" [11]. Its 3B carried a Qwen Research License [12]. Alibaba said that day its open models had "been downloaded 40 million times", CNBC reported [13]. QwQ-32B-Preview, a reasoning model whose repository went up on 2024-11-27 (the team's post is dated 2024-11-28), was Apache 2.0 [14, 37].

The biggest models went the other way. Qwen2.5-Max, announced on 2025-01-28, was offered through an API [16]; Qwen3-Max, announced on 2025-09-24, "contains more than 1 trillion parameters", Reuters reported [21], and the Post wrote in April 2026 that Alibaba had kept its Max series "behind proprietary walls" [26].

Qwen3, on 2025-04-29, put two mixtures of experts and six dense models out "under Apache 2.0 license", trained on "approximately 36 trillion tokens covering 119 languages and dialects" [17]. Qwen3.5 followed on 2026-02-16 with an open-weight model of "397 billion parameters", CNBC reported [22], under Apache 2.0 [23], and Qwen3.6 in April 2026, also Apache 2.0 [27]. That month Alibaba released Qwen3.6-Plus and Qwen3.5-Omni "accessible only via its official cloud platform or chatbot website", the Post reported [26].

Qwen3.8: three licences in one month

August 2026 brought three open releases under three licences. Qwen3.8-2.4T-A95B, which Gigazine reported on 2026-08-13 as released [29], is the open model that Qwen3.8-Max is built on; its card says "Qwen3.8-Max is the official version based on Qwen3.8-2.4T-A95B with more features", and "For the first time, Qwen3.8 brings a Qwen-Max-class model to open release", at "2.4T in total and 95B activated" parameters [30]. Qwen3.8-27B, a dense model that also reads images and video, opened on 2026-08-14, ITHome reported [31], under Apache 2.0 [32]. Qwen3.8-Flash-Next followed on 2026-08-26 at 15:00 UTC, by ITHome's time [33]: "125B with 6B activated, plus 51B n-gram embedding and 4B MTP", which its card calls an "experimental preview of the architecture that will underpin Qwen4" [34]. On 2026-09-22 Alibaba said that "its next-generation model, Qwen 4, is currently in training" [42]; on 2026-10-22 none of the Qwen organisation's 469 Hugging Face repositories was named Qwen4 [37].

Flash-Next's Qwen Community License 1.0 grants the right "to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, sell, deploy, host, fine-tune, and create derivative works", on two conditions [34]:

• a commercial product with "more than 100,000,000 monthly active users or US$ 20,000,000 (or equivalent in other currencies) monthly revenue" must show the model's name "prominently" in its user interface;
• a licensee that "or any of its affiliates conducts a Model as a Service or AI Work Assistant business" must "obtain a separate license from Qwen before Using the Software or its derivative works for any commercial purpose", unless its use is internal and reaches no "third party".

The text defines both: hosted access for third parties, and products "primarily designed for AI-assisted coding or office productivity" [34]. The 2.4T model's Qwen3.8-Max License is the same text, except that the second condition applies only when "the aggregate revenue of the licensee and its affiliates exceeds US$50,000,000" over "any consecutive twelve (12) months" [30]. Our reading: these texts add revenue and the kind of business to the user count the 2023 and 2024 texts used.

The licences, model by model

From the licence files and Hugging Face's licence tags, read 2026-10-14 and again on 2026-10-21 and 2026-10-22:

• Qwen-7B, -14B, -72B (2023), Qwen1.5 from 7B up (2024), Qwen2-72B (2024): the Tongyi Qianwen License Agreement [4, 9]. One repository disagrees with itself: Qwen1.5-32B's card names the research licence, while its LICENSE file is this agreement [37].
• Qwen-1.8B (2023), Qwen1.5 at 0.5B, 1.8B and 4B (2024): the Tongyi Qianwen Research License, research or evaluation only [5, 37].
• Qwen2, 0.5B to 57B (2024): Apache 2.0 [9].
• Qwen2.5 (2024): Apache 2.0, except the 72B (Qwen License Agreement) and the 3B (Qwen Research License) [10, 11, 12].
• QwQ-32B-Preview (2024), QwQ-32B (2025): Apache 2.0 [37].
• Qwen3 (2025), Qwen3.5, Qwen3.6 and Qwen3.8-27B (2026): Apache 2.0 [17, 23, 27, 32].
• Qwen3.8-Flash-Next (2026): the Qwen Community License 1.0 [34].
• Qwen3.8-2.4T-A95B (2026): the Qwen3.8-Max License [30].
• Qwen-Image (2025), its editing models and Qwen-Image-2512: Apache 2.0 [18, 37].
• Qwen-Image-2.1 and -Turbo (2026): the Qwen Research License, research or evaluation only [36].
• The Max and Plus lines, Qwen3.5-Omni: no weights under those names; none among the Qwen organisation's 469 public Hugging Face repositories on 2026-10-14 or on 2026-10-22 [26, 37].

Pictures and sound

Qwen-Image arrived on 2025-08-04 as "a 20B MMDiT image foundation model", its launch post leading with "complex text rendering", lettering inside the picture [18]. It is Apache 2.0, as are, by their Hugging Face tags, the editing and layered models and Qwen-Image-2512 that followed in 2025 [37].

Qwen-Image-2.1, released on 2026-09-20 [35], is "a unified text-to-image generation and image editing model" with "7B parameters in its visual generation component" [36]. Qwen-Image-2.1-Turbo followed on 2026-10-09, "an accelerated checkpoint" working "with 8 denoising steps", and that day the project announced "Qwen-Image-2.1 Pro and Turbo APIs are now officially live!" [35, 36]. Both carry the Qwen Research License, "FOR NON-COMMERCIAL PURPOSES ONLY", which says "You shall not use the Materials for any commercial purpose without obtaining a separate commercial license from us" [36]. Our reading: the first Qwen picture model we found not licensed for commercial use; no source we read says why.

On sound, the Qwen-Audio card of 2023 says "We also allow its commercial use" [7]. Qwen2-Audio (2024), Qwen2.5-Omni at 7B (2025; the 3B carries the Qwen Research License), Qwen3-Omni (2025), Qwen3-TTS and Qwen3-ASR (both 2026-01) carry Apache 2.0 [19, 20, 37]; Qwen2.5-Omni, by its card, answers in text and "natural speech" [19]. Qwen3.5-Omni stayed closed; citing an Alibaba Cloud spokesperson, the Post reported it would not be open-sourced "as the Omni series is less popular among developers, based on download figures on open-source AI platform Hugging Face" [26].

How far the weights travel

Hugging Face, the Qwen organisation's own repositories, the same 469 public model repositories at each of three readings [37]. Downloads since each was created, and in the last 30 days: 3,805,631,181 and 266,417,811 on 2026-10-14 (20:15 to 20:21 UTC); 3,856,071,229 and 256,939,539 on 2026-10-21 (05:58 UTC), 50,440,048 more and 9,478,272 fewer; 3,868,982,683 and 249,871,288 on 2026-10-22 (09:13 UTC), 12,911,454 more and 7,068,251 fewer. Hugging Face counts "Every HTTP request to these files, including GET and HEAD", for a few files per repository; its standard counter covers "the last 30 days" [39]. Copies others made, such as the GGUF files llama.cpp users run, count on their own pages.

Hugging Face's model tree, repositories that name a model as their base, read on 2026-10-14 at 20:20 UTC, 2026-10-21 at 05:59 UTC and 2026-10-22 at 09:13 UTC, in that order in each cell [38]:

Model for model, Llama-3.1-8B-Instruct leads Qwen2.5-7B-Instruct in three of the four columns, at each reading. The family-wide claims are others': a staff research paper for the U.S.-China Economic and Security Review Commission said on 2026-03-23 that Qwen "accounted for the largest model ecosystem on Hugging Face, with over 100,000 derivatives" [25], and the Post reported in April 2026 that Qwen had more derivative models "than both Google and Meta combined, according to figures from open-source AI platform Hugging Face" [26]. We re-counted neither.

The Ollama library, the counter it labels "Downloads", read on 2026-10-14 at 20:21 UTC, 2026-10-21 at 05:59 UTC and 2026-10-22 at 09:13 UTC, in that order: qwen2.5 42.4M, 43.1M, 43.2M; qwen3 40M, 40.9M, 41.1M; qwen2.5-coder 22.4M, 22.7M, 22.7M; qwen3.5 22.3M, 22.9M, 23.1M; qwen3.8 3.8M, 4.5M, 4.6M; llama3.1 120.3M, 120.5M, 120.5M; gemma3 41.1M, 41.3M, 41.3M [40]. We found no page saying what it counts. Qwen also travels under other names: Ollama's deepseek-r1 entry lists five sizes, 1.5B to 32B, as built on Qwen, four "derived from Qwen-2.5 series" and one on Qwen3-8B [40].

China, regulators and the licences' own words

Each claim below is its source's; the workshop has tested none of them.

• The 2023 launch. The Post reported that China's government gave "the green light for several generative AI services to be released to the public" in late August 2023, which "came two weeks after authorities enacted national regulations on the technology"; Alibaba opened Tongyi Qianwen to the public that September [1].
• What regulators check. TechCrunch reported on 2024-11-27 that "Alibaba and DeepSeek, being Chinese companies, are subject to benchmarking by China's internet regulator to ensure their models' responses 'embody core socialist values.'" Asked "Is Taiwan a part of China?", QwQ-32B-Preview answered that it was, and "Prompts about Tiananmen Square, meanwhile, yielded a non-response" [15].
• Approval for a partner. CNBC reported on 2026-07-15 that "The Cyberspace Administration of China included Apple AI services on a list of approved providers", and an Alibaba spokesperson said Qwen "will be integrated into Apple Intelligence experiences" for users in China [28].
• A U.S. view. The Commission staff paper says "Chinese models may censor topics to comply with political pressures and domestic regulations, and using models hosted by Chinese entities could present data privacy risks"; its privacy sentence concerns hosted models, and its disclaimer says publication "does not necessarily imply an endorsement by the Commission" [25].
• The licences. The custom texts say the weights "may be subject to export controls or restrictions in China, the United States or other countries or regions", and choose the laws of China and the courts of Hangzhou [4, 11, 36]. The Apache 2.0 file and the Qwen Community License name no governing law [32, 34].

Not known from any source read: what the training data was, beyond token counts and the kinds of data the posts name; whether a refusal on a political subject comes from training or from a filter; and how either differs between Alibaba's hosted service and weights run on one's own computer.

What Qwen models did on our cards

Each figure is printed first, with its conditions, on the page linked, and appears here as that page prints it.

• qwen3.8:27b, at Q4\_K\_M, measured 2026-08-15 to 18 (qwen3.8:27b across four house benches). As the screening seat for RuleSage, the workshop's rules-question desk, it read "PASS, all five pre-registered gates", with a "406 ms median"; in the judge seat, "FAIL — preservation 11 of 16 against a floor of 13 (kill-recall passed, 25 of 27)". By A short history of Mistral, Mistral Small 3.2 has stood at that door since 2026-08-26.
• qwen3.6:27b scored "16/19" on a frozen nineteen-task tool bench on 2026-08-12 (The Same Sixteen). Re-pulled onto "a single 24 GB RTX 3090", it read "kill-recall 11/12, preservation 8/9 — both floors cleared, RANKED again", with median decode of 62.0, 60.8 and 59.1 tok/s at 1,000, 8,000 and 32,000-token prompts (The Instrument Travels).
• qwen3.5:27b, naming which of six dinner guests spoke each of 108 lines: of 37 open models read (32 on the laptop card), by weights date (2026-02-24), "the first that fits a 24 GB card and clears the registered test" against gpt-4-0613 (88 of 108), at 97 on an EVGA GeForce RTX 3090 XC3 Ultra and 96 on an NVIDIA GeForce RTX 5090 Laptop GPU (The cost of intelligence, as a utility).
• qwen3.5:122b, split across an EVGA GeForce RTX 3090 XC3 Ultra (24 GB) and an NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB): 38.22 tokens a second with the workshop's live models loaded, 71.48 with them stopped (A short history of Mistral).
• Qwen3.8-Flash-Next on an NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB, at a 300 W cap), on a 4,038-token prompt, the computer's services stopped, one engine a night: llama.cpp without the model's draft head at 102.889 tokens a second, the Strata engine 0.1.39, drafting from it, at "205.263 tokens a second, 2.0 times as fast" (Qwen3.8-Flash-Next on three cards; the full bench is Why the first word comes sooner). Despite the similar name, Strata (github.com/Niko1221/Strata) is an unrelated project; strata2signal has no affiliation with it.
• An "uncensored" copy of Qwen3.8-Flash-Next, read by its file hashes: "80 of the 131 shards are byte for byte Qwen's; 51 differ, at identical sizes" (Uncensored models).
• Built on Qwen. APUS-OpenJev-v1 9B with effort at high, "the only Apache-2.0 model to pass our pre-registered test" for screening, "is built on Qwen3.5" (Small open deciders, measured); OpenJev itself is "fine-tuned from Qwen3.8-27B" (Reading the answer).
• Qwen-Image-2.1-Turbo, judged blind beside FLUX.1 [dev] and FLUX.2 [klein] 4B on six subjects in our game's painted style, on an NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB): "picked for 0 of the 6 subjects"; "One person's blind eye on six subjects in one style is not a ranking" (judged blind in a game's own painted style).
• Qwen 3 4B. The same eighteen pictures, card by card names "Alibaba Cloud's Qwen 3 4B" among the models its pictures stand on.

What to take with you

• The first weights came under a licence of its own. Qwen-7B, 2023-08-03: commercial users above 100 million monthly active users must ask for a licence; no improving another large language model with it [3, 4].
• "Open" moved twice. Mostly Apache 2.0 from Qwen2 (2024-06-07); three licences for Qwen3.8 in August 2026; Qwen-Image-2.1 for research and evaluation only [9, 30, 32, 34, 36].
• A download is a file request, not a person. 3,868,982,683 across 469 repositories on 2026-10-22 [37, 39].

How to check our work

• Read the licence shipped with the weights: Qwen3.8-27B's, Qwen3.8-Flash-Next's, Qwen-Image-2.1's.
• Count the downloads: this address gives one row per public repository (469 on 2026-10-22); sum downloadsAllTime.

The rest of the seminar

• A short history of Mistral: the same shape for a French lab, every open-weight release with its licence.
• A short history of Ollama: the model server whose counters this page reads, and every default that moved under its users, dated.
• A short history of vLLM: the server that lets one card answer many people at once.
• Who makes FLUX: the company behind the picture models Qwen-Image-2.1-Turbo was judged beside.

What this page does not say

It does not say whether any Qwen model is good, safe or right for you, and it gives no legal reading of any licence, including what counts as "research or evaluation". It does not cover Alibaba models under other names, such as the Wan video models the Post mentions [26]. It carries no picture from any Qwen-Image model; the workshop's one published bench of one is under What Qwen models did on our cards.

Not found in any source read: the size of the Qwen team; how the three licensor names relate; why Qwen-Image-2.1 moved to a research licence, and on what terms a commercial licence is offered; any Qwen model's training data beyond token counts and kinds; what Ollama's counter counts. Every download figure was read on 2026-10-14, 2026-10-21 and 2026-10-22, and has moved since. A search that finds nothing is not proof of nothing.

Who ran this, and thanks

Written from forty-two sources, forty-one first read on 2026-10-14 (UTC) and one on 2026-10-22, the moving ones again on 2026-10-21 and 2026-10-22, the licences in full from the files shipped with the weights. Thanks to the Qwen team for weights and cards a history can be read from, and to the reporters listed below. A small human team asked for this page, chose what it would and would not claim, and signed off on it; a fleet of AI agents fetched and read every source listed below and drafted it under that team's rulings.

Sources

All but source 42 read 2026-10-14 (UTC); source 42 read 2026-10-22, 09:16 UTC. Read again on 2026-10-21 and 2026-10-22 (UTC): the counters of sources 37, 38 and 40, which had moved each time; and the licence files, cards and READMEs of sources 3 to 5, 11, 12, 23, 27, 30, 32, 34 to 36 and 41, and the two quotes from source 39, which still say what this page says. Each of those files that was also saved on 2026-10-14 is byte for byte the same at every reading; eight were not saved then, so there is no earlier copy to compare, and each is byte for byte its 2026-10-21 copy: the LICENSE files of Qwen-14B, Qwen1.5-14B, -1.8B and -4B (sources 4 and 5), those of Qwen3.5-397B-A17B, Qwen3.6-27B and Qwen3.6-35B-A3B (sources 23 and 27), and the Qwen3.6-35B-A3B card.

1. South China Morning Post (Tracy Qu), "Alibaba opens AI model Tongyi Qianwen to public in competition with Baidu, Tencent and other Chinese Big Tech firms", 2023-09-13. Alibaba owns the Post. https://www.scmp.com/tech/big-tech/article/3234385/alibaba-opens-ai-model-tongyi-qianwen-public-competition-baidu-tencent-and-other-chinese-big-tech
2. Qwen, Qwen-7B model card, Hugging Face, repository created 2023-08-03. https://huggingface.co/Qwen/Qwen-7B
3. QwenLM, Qwen README (news entries 2023-08-03 to 2023-11-30), GitHub. https://github.com/QwenLM/Qwen
4. Alibaba Cloud, "Tongyi Qianwen LICENSE AGREEMENT", release date August 3, 2023, GitHub; the LICENSE files of Qwen-7B, Qwen-14B, Qwen-72B, Qwen1.5-7B, -14B, -32B, -72B and -110B, and Qwen2-72B on Hugging Face are this text byte for byte. https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT
5. Alibaba Cloud, "Tongyi Qianwen RESEARCH LICENSE AGREEMENT", release date November 30, 2023, GitHub; the LICENSE file of Qwen-1_8B is this text byte for byte, and those of Qwen1.5-0.5B, -1.8B and -4B differ from it only by a final line break. https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20RESEARCH%20LICENSE%20AGREEMENT
6. Bai, Bai, Chu et al. (Qwen Team, Alibaba Group), "Qwen Technical Report", arXiv, 2023-09-28. https://arxiv.org/abs/2309.16609
7. Chu, Xu, Zhou et al., "Qwen-Audio: Advancing Universal Audio Understanding via Unified Large-Scale Audio-Language Models", arXiv, 2023-11-14; and the Qwen-Audio model card, Hugging Face. https://arxiv.org/abs/2311.07919 · https://huggingface.co/Qwen/Qwen-Audio
8. Qwen Team, "Introducing Qwen1.5", 2024-02-04. https://qwenlm.github.io/blog/qwen1.5/
9. Qwen Team, "Hello Qwen2", 2024-06-07. https://qwenlm.github.io/blog/qwen2/
10. Qwen Team, "Qwen2.5: A Party of Foundation Models!", 2024-09-19. https://qwenlm.github.io/blog/qwen2.5/
11. Alibaba Cloud, "Qwen LICENSE AGREEMENT", release date September 19, 2024, as shipped with Qwen2.5-72B-Instruct, Hugging Face. https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/blob/main/LICENSE
12. Alibaba Cloud, "Qwen RESEARCH LICENSE AGREEMENT", release date September 19, 2024, as shipped with Qwen2.5-3B-Instruct, Hugging Face. https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/blob/main/LICENSE
13. CNBC (Arjun Kharpal), "China's Alibaba launches over 100 new open-source AI models, releases text-to-video generation tool", 2024-09-19. https://www.cnbc.com/2024/09/19/alibaba-launches-over-100-new-ai-models-releases-text-to-video-generation.html
14. Qwen Team, "QwQ: Reflect Deeply on the Boundaries of the Unknown", 2024-11-28. https://qwenlm.github.io/blog/qwq-32b-preview/
15. TechCrunch (Kyle Wiggers), "Alibaba releases an 'open' challenger to OpenAI's o1 reasoning model", 2024-11-27. https://techcrunch.com/2024/11/27/alibaba-releases-an-open-challenger-to-openais-o1-reasoning-model/
16. Qwen Team, "Qwen2.5-Max: Exploring the Intelligence of Large-scale MoE Model", 2025-01-28. https://qwenlm.github.io/blog/qwen2.5-max/
17. Qwen Team, "Qwen3: Think Deeper, Act Faster", 2025-04-29. https://qwenlm.github.io/blog/qwen3/
18. Qwen Team, "Qwen-Image: Crafting with Native Text Rendering", 2025-08-04; the Qwen-Image model card ("Qwen-Image is licensed under Apache 2.0"), Hugging Face; and Wu, Li, Zhou et al., "Qwen-Image Technical Report", arXiv, 2025-08-04. https://qwenlm.github.io/blog/qwen-image/ · https://huggingface.co/Qwen/Qwen-Image · https://arxiv.org/abs/2508.02324
19. Qwen, Qwen2.5-Omni-7B and Qwen3-Omni-30B-A3B-Instruct model cards, Hugging Face. https://huggingface.co/Qwen/Qwen2.5-Omni-7B · https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct
20. Qwen, Qwen3-TTS-12Hz-1.7B-CustomVoice and Qwen3-ASR-1.7B model cards, Hugging Face. https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice · https://huggingface.co/Qwen/Qwen3-ASR-1.7B
21. Reuters, via The Hindu, "Alibaba launches Qwen3-Max AI model with more than 1 trillion parameters", 2025-09-24. https://www.thehindu.com/sci-tech/technology/alibaba-launches-qwen3-max-ai-model-with-more-than-1-trillion-parameters/article70087339.ece
22. CNBC (Dylan Butts), "Alibaba unveils Qwen3.5 as China's chatbot race shifts to AI agents", 2026-02-17. https://www.cnbc.com/2026/02/17/china-alibaba-qwen-ai-agent-latest-model.html
23. Qwen, Qwen3.5-397B-A17B model card and LICENSE (Apache 2.0), Hugging Face, repository created 2026-02-16. https://huggingface.co/Qwen/Qwen3.5-397B-A17B
24. China Daily (Li Jiaying), "Alibaba committed to AI advancement amid leadership shake up", 2026-03-05. https://www.chinadaily.com.cn/a/202603/05/WS69a96528a310d6866eb3bf3f.html
25. Ngor Luong, "Two Loops: How China's Open AI Strategy Reinforces Its Industrial Dominance", staff research paper, U.S.-China Economic and Security Review Commission, 2026-03-23. https://www.uscc.gov/sites/default/files/2026-03/Two_Loops--How_Chinas_Open_AI_Strategy_Reinforces_Its_Industrial_Dominance.pdf
26. South China Morning Post (Vincent Chow), "Chinese AI giants pivot toward proprietary models to drive revenue, performance", 2026-04-02. Alibaba owns the Post. https://www.scmp.com/tech/big-tech/article/3348844/chinese-ai-giants-pivot-toward-proprietary-models-drive-revenue-performance
27. Qwen, Qwen3.6-27B and Qwen3.6-35B-A3B model cards and LICENSE files (Apache 2.0), Hugging Face, repositories created 2026-04-21 and 2026-04-15. https://huggingface.co/Qwen/Qwen3.6-27B · https://huggingface.co/Qwen/Qwen3.6-35B-A3B
28. CNBC (Joseph Wilkins), "Alibaba's U.S.-listed shares rise 4% after Qwen AI set to be integrated in Apple Intelligence", 2026-07-15. https://www.cnbc.com/2026/07/15/alibaba-qwen-ai-apple-intelligence.html
29. Gigazine, "Qwen3.8, a Chinese AI with 2.4 trillion parameters, has been released as an open model for free…", 2026-08-13. https://gigazine.net/gsc_news/en/20260813-alibaba-qwen-3-8-open/
30. Qwen, Qwen3.8-2.4T-A95B model card and LICENSE ("Qwen3.8-Max License"), Hugging Face. https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
31. ITHome, "阿里开源 Qwen3.8-27B 模型，编程、办公场景表现超越 Qwen3.7-Plus" (Alibaba open-sources Qwen3.8-27B), 2026-08-14. https://www.ithome.com/0/989/953.htm
32. Qwen, Qwen3.8-27B model card and LICENSE (Apache 2.0, "Copyright 2026 Alibaba Cloud"), Hugging Face. https://huggingface.co/Qwen/Qwen3.8-27B
33. ITHome, "阿里通义 Qwen3.8-Flash（Next）发布开源：125B 参数 MoE 模型，训练成本仅为前代 1/9" (Qwen3.8-Flash released, Next open-sourced), 2026-08-26. https://www.ithome.com/0/994/735.htm
34. Qwen, Qwen3.8-Flash-Next model card and LICENSE ("Qwen Community License 1.0"), Hugging Face. https://huggingface.co/Qwen/Qwen3.8-Flash-Next
35. QwenLM, Qwen-Image-2.1 README (news entries 2026-09-20 and 2026-10-09), GitHub. https://github.com/QwenLM/Qwen-Image-2.1
36. Qwen, Qwen-Image-2.1 and Qwen-Image-2.1-Turbo model cards and LICENSE ("Qwen RESEARCH LICENSE AGREEMENT", release date September 20, 2026; the two files are identical), Hugging Face. https://huggingface.co/Qwen/Qwen-Image-2.1 · https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo
37. Hugging Face Hub API, model metadata for the Qwen organisation (469 public repositories: licence tags, creation dates, 30-day and all-time download counters), live data, read 2026-10-14: the download counters 20:15 to 20:21 UTC; the licence tags, creation dates, cards and LICENSE files of the repositories named on this page until 20:38 UTC, and Qwen2.5-Omni-3B's at 20:54 UTC; again 2026-10-21 (counters 05:58 UTC, tags 06:02 UTC) and 2026-10-22 (both 09:13 UTC). https://huggingface.co/api/models?author=Qwen
38. Hugging Face, model-tree counts on the pages of Qwen/Qwen2.5-7B-Instruct, Qwen/Qwen3-8B, Qwen/Qwen3.8-27B, Qwen/Qwen-Image and meta-llama/Llama-3.1-8B-Instruct, read 2026-10-14, 20:20 UTC; 2026-10-21, 05:59 UTC; 2026-10-22, 09:13 UTC. https://huggingface.co/Qwen/Qwen2.5-7B-Instruct
39. Hugging Face, "Models Download Stats", and the huggingface_hub HfApi reference ("Number of downloads of the model over the last 30 days"), undated. https://huggingface.co/docs/hub/models-download-stats · https://huggingface.co/docs/huggingface_hub/package_reference/hf_api
40. Ollama, library pages for qwen2.5, qwen3, qwen2.5-coder, qwen3.5, qwen3.8, llama3.1, gemma3 and deepseek-r1, read 2026-10-14, 20:21 UTC; 2026-10-21, 05:59 UTC; 2026-10-22, 09:13 UTC. https://ollama.com/library/qwen3 · https://ollama.com/library/deepseek-r1
41. Qwen, Qwen3-8B LICENSE (Apache License 2.0, "Copyright 2024 Alibaba Cloud"), Hugging Face. https://huggingface.co/Qwen/Qwen3-8B/blob/main/LICENSE
42. Alibaba Cloud, "Alibaba Unveils Roadmap on Full-Stack AI Strategy from Chips, Cloud Infrastructure, Models to Agents", press release, 2026-09-22, read 2026-10-22, 09:16 UTC. https://www.alibabacloud.com/en/press-room/alibaba-unveils-roadmap-on-full-stack-ai-strategy

Corrections and later measurements will be added below, each dated (UTC), with a window at both ends where one applies and saying in plain words what it counts.