// Chapter 01 · Ecosystem

Each Source Has a Different Leader

Qwen takes more than half of the downloads we track on Hugging Face, mostly through small models. We assume they run on home GPUs. But is it what companies use?

Data: Hugging Face download and repository statistics (cfahlgren1/hub-stats, Apache-2.0), ATOM Project tracked-model list, OpenAlex (CC0).

54%

of the downloads of the models we track on Hugging Face from October 2025 to September 2026 went to Qwen.

Hugging Face downloads (hub-stats); ATOM Project.

Illustration: layered landscape with a pale sun
Downloads, derivatives, mentions

From October 2025 to September 2026 Qwen accounted for 54% of the downloads of the open models we track on Hugging Face, and eight of the ten most downloaded repositories are its own. Most of those downloads are of small models, which may be easier to distribute and to try on home hardware. But a model that is downloaded a lot is not necessarily the one used most: in the tokens served through OpenRouter, DeepSeek comes first.

01 Most downloaded

Qwen has eight of the year’s ten most downloaded models

The ranking is led by small models. The most downloaded, Qwen3-0.6B, has 0.6 billion parameters; twelve of the fifteen have fewer than 10 billion.

The most downloaded open models in our selection on Hugging Face

Qwen Other developers

October 2025 to September 2026 · share of all downloads of the models we track

  1. 1

    Qwen3-0.6B Qwen

    201.9 million 5.1%

  2. 2

    Qwen2.5-7B-Instruct Qwen

    137.5 million 3.5%

  3. 3

    Qwen3-8B Qwen

    107.5 million 2.7%

  4. 4

    Qwen2.5-1.5B-Instruct Qwen

    103.5 million 2.6%

  5. 5

    Llama-3.1-8B-Instruct Meta

    96.6 million 2.4%

  6. 6

    Qwen2.5-3B-Instruct Qwen

    92.9 million 2.3%

  7. 7

    gpt-oss-20b OpenAI

    81.1 million 2.0%

  8. 8

    Qwen2.5-VL-3B-Instruct Qwen

    72.6 million 1.8%

  9. 9

    Qwen3-4B Qwen

    70.6 million 1.8%

  10. 10

    Qwen2.5-VL-7B-Instruct Qwen

    68.8 million 1.7%

Show places 11 to 15
  1. 11

    Llama-3.2-1B-Instruct Meta

    68.0 million 1.7%

  2. 12

    Qwen3-VL-8B-Instruct Qwen

    67.0 million 1.7%

  3. 13

    Qwen3.5-9B Qwen

    62.7 million 1.6%

  4. 14

    gemma-4-26B-A4B-it Google

    62.0 million 1.6%

  5. 15

    gemma-4-31B-it Google

    58.7 million 1.5%

Downloads of the open models on the ATOM Project’s tracked list, summed over Hugging Face’s 30-day readings. A download is a request for a repository’s files, not a user. Source: Hugging Face hub-stats; ATOM Project.

Without Qwen: who leads the downloads

Leaving Qwen out: the most downloaded families and models, last 12 months

  1. 1

    Gemma 4 Google

    201.7 million 5.1%

  2. 2

    gpt-oss OpenAI

    131.1 million 3.3%

  3. 3

    Llama 3.2 Meta

    128.1 million 3.2%

  4. 4

    Llama 3.1 Meta

    124.4 million 3.1%

  5. 5

    DeepSeek-R1 DeepSeek

    91.3 million 2.3%

  6. 6

    Gemma 3 Google

    90.1 million 2.3%

  7. 7

    Mistral 7B Mistral AI

    70.0 million 1.8%

  8. 8

    Nemotron 3 NVIDIA

    63.4 million 1.6%

The same downloads and period without Qwen’s models. Shares are of all the downloads, Qwen included, so they are comparable with the ranking above. A family groups one developer’s models of one line or generation. Source: Hugging Face hub-stats; ATOM Project.

Without Qwen, the most downloaded families of the year come from US labs: Gemma 4 (Google), gpt-oss (OpenAI) and Llama 3.2 (Meta). None comes close: Gemma 4, the largest, took 5% of all the downloads, about a tenth of Qwen’s 54%, and Llama 3.2 and Llama 3.1 together 6%.

02 Downloads

The labs behind the most downloaded models have changed

Since August 2024 the balance between labs has shifted. Almost all of the most downloaded open models now come from the United States or China. But since August 2025 models from developers based in China have taken the larger share every month. In October 2026 they took 68% of the downloads.

Share of downloads of models developed in China

68%

in October 2026, against 17% in August 2024. Each square is one percentage point.

Figure 1.1

Share of downloads of tracked open models, by developer region

From August 2025, developers based in China have led those based in the US every month.

  • China-based 68%
  • of which Qwen 55%
  • US-based 30%
  • Europe-based 2%
  • China-based, without Qwen 29%
  • China-based, models present in Aug 2024 23%
Share of observed open-model downloads by developer region, with Qwen's partMonthly, 2024-08 to 2026-10. China-based developers' share of observed downloads of the tracked models: 17% at the start (24% without the outlier cap) and 68% in 2026-10. Qwen alone: 55% of all downloads in 2026-10. US-based 69% to 30%; Europe-based 12% to 2%. Without Qwen the China-based share goes from 9% to 29%; among the 486 models already present in 2024-08, 23%. Downloads, not users or deployments.0%25%50%75%100%Aug 2024Jan 2025Jul 2025Jan 2026Jul 2026Oct 2026count changeQwen’s models
Share of observed open-model downloads by developer region, with Qwen's partMonthly, 2024-08 to 2026-10. China-based developers' share of observed downloads of the tracked models: 17% at the start (24% without the outlier cap) and 68% in 2026-10. Qwen alone: 55% of all downloads in 2026-10. US-based 69% to 30%; Europe-based 12% to 2%. Without Qwen the China-based share goes from 9% to 29%; among the 486 models already present in 2024-08, 23%. Downloads, not users or deployments.0%25%50%75%100%Aug ’24Jan ’26Oct ’26count changeQwen

Source Hugging Face Hub download counters, via the cfahlgren1/hub-stats archive; ATOM Project model list (Apache-2.0)

Qwen explains much of the change, but not all of it. Leave Qwen’s models out and the other developers based in China still rise, from 9% to 29% of the remaining downloads, mostly through DeepSeek and Z.ai. Without Qwen, though, US-based labs would still lead: in October 2026 they took 66% of the downloads that did not go to Qwen.

How we measured it
Share of 30-day downloads, read on the first of each month; each model is capped at five times its own median month. Region is the developer’s headquarters as labelled by ATOM, not where downloads come from.

Without the cap the first month starts at 24%, not 17%. Grey band, December 2024 to February 2025: counting jump (Mistral), not interpreted.

Alternative populations, not a margin of error, in the data table: without Qwen’s models, and the 486 models that existed in August 2024.

Source
Hugging Face Hub download counters, via the cfahlgren1/hub-stats archive (archive licensed Apache-2.0); model list and region labels: ATOM Project tracked-models @d4b1709 (Apache-2.0)
Base
The 1,676 tracked models (ATOM list of models with published weights) with downloads on 1 October 2026
Evidence
E3our tabulation
Figure 1.1 data: shares of observed downloads by developer region
MonthModelsChina-basedof which QwenUS-basedEurope-basedChina, uncappedChina without QwenChina, 486 models of Aug 2024
2024-0844117.2%9.5%68.8%11.8%23.7%8.5%17.2%
2024-0947519.7%13.1%65.7%12.7%24.7%7.6%19.4%
2024-1055026.1%15.2%64.4%8.2%30.7%12.9%26.4%
2024-1158147.0%42.8%46.7%5.5%47.5%7.2%20.4%
2024-1263249.1%45.1%46.4%3.8%55.4%7.2%25.1%
2025-0167424.5%19.6%45.2%29.4%14.9%6.1%15.8%
2025-0273227.0%18.3%50.8%21.0%24.1%10.6%17.9%
2025-0376340.5%19.2%50.3%8.5%39.8%26.4%20.0%
2025-0480341.9%23.0%48.9%8.4%41.9%24.5%22.2%
2025-0589045.0%28.5%48.0%6.1%44.8%23.1%21.9%
2025-0693845.7%32.4%47.9%5.6%46.9%19.7%17.0%
2025-0793946.4%32.6%47.3%5.5%47.4%20.5%17.1%
2025-08109657.3%45.7%38.5%3.8%61.8%21.3%29.6%
2025-09116353.1%40.3%44.1%2.4%49.7%21.6%28.9%
2025-10121156.2%44.3%41.1%2.4%56.1%21.4%23.7%
2025-11128160.0%47.8%35.6%4.2%60.5%23.4%20.8%
2025-12130763.8%50.0%32.3%3.8%64.7%27.7%21.5%
2026-01138261.4%47.3%34.3%4.1%62.2%26.8%22.3%
2026-02139066.2%55.6%30.6%3.1%67.2%23.8%23.2%
2026-03142068.2%56.3%28.0%3.7%69.0%27.1%21.9%
2026-04146467.5%56.0%29.5%2.9%67.6%26.0%22.7%
2026-05152167.6%55.0%29.2%3.1%77.4%27.9%24.2%
2026-06153966.8%54.3%30.0%3.0%70.4%27.4%26.0%
2026-07158066.3%53.8%31.3%2.2%65.8%27.0%30.6%
2026-08159965.0%53.7%32.3%2.5%64.8%24.5%25.3%
2026-09164368.6%57.3%29.2%2.0%68.0%26.5%23.3%
2026-10167668.2%55.4%29.6%2.0%67.0%28.7%23.4%

03 Qwen's models

Which Qwen models are downloaded most

All ten of Qwen’s most downloaded models over the year have fewer than 10 billion parameters. Three of them show the range: a tiny text model, a mid-size model tuned for instructions and one that also reads images.

Three of Qwen’s most downloaded models, last 12 months

  1. Qwen/Qwen3-0.6B

    A small text model

    0.6 billion parametersQwen3#1 overall

    201.9 million downloads

  2. Qwen/Qwen2.5-7B-Instruct

    A mid-size model tuned for instructions

    7.61 billion parametersQwen2.5#2 overall

    137.5 million downloads

  3. Qwen/Qwen2.5-VL-3B-Instruct

    A model that also reads images

    3 billion parametersQwen2.5#8 overall

    72.6 million downloads

Downloads summed over twelve 30-day readings, each model capped at five times its own median month; the rank is among all the models we track. Parameters as published by Qwen. Source: Hugging Face hub-stats; ATOM Project.

Two generations carry the family: among Qwen’s models in the year’s top 100, Qwen3 (2025) took 44% of their downloads and Qwen2.5 (2024) 35%. Older models keep being fetched long after newer ones arrive.

04 Downloaded and used

The most downloaded are not the most used

From October 2025 to September 2026 Qwen took 54% of the downloads we track; in the same months it took 1.0% of the open-weight tokens in OpenRouter’s daily top 50, where DeepSeek led with 31%. Downloads and use do not always go together.

Each source has a different leader: Qwen in downloads, DeepSeek in tokens and Meta in requests

The first five labs of each source, each on its own scale. A lab has the same colour in all three columns; hover over one to see it in the others.

DownloadsHugging Face, models we trackOct. 2025 – Sept. 2026

  1. 1Qwen54.2% of downloads
  2. 2Google8.5% of downloads
  3. 3Meta8.2% of downloads
  4. 4DeepSeek6.1% of downloads
  5. 5OpenAI3.3% of downloads

TokensOpenRouter, open weights in the daily top 50Oct. 2025 – Sept. 2026

  1. 1DeepSeek31.3% of tokens
  2. 2Tencent16.5% of tokens
  3. 3Xiaomi12.2% of tokens
  4. 4Z.ai12.1% of tokens
  5. 5MiniMax7.3% of tokens

RequestsCloudflare Workers AI, text generationOct. 2026

  1. 1Meta29.1% of requests
  2. 2Z.ai17.9% of requests
  3. 3DeepSeek11.0% of requests
  4. 4Qwen8.0% of requests
  5. 5OpenAI7.0% of requests
Grey: labs that are among the first five of one source only. A download is a request for files, not a user; the tokens are those of versions verified as open-weight; requests are not tokens. The shares are not market shares and cannot be added. Sources: Hugging Face hub-stats and the ATOM Project; OpenRouter (CC BY 4.0); Cloudflare Radar (used with permission).

Each measure crowns a different lab: Qwen leads the downloads (54%), DeepSeek the tokens (31%) and Meta the requests on Workers AI (29%). Only DeepSeek is among the four leaders of all three. There is no single answer to which open model is used most: downloads, API traffic and a platform like Cloudflare each see different uses.

05 What gets built on top

What is being built on top of Qwen

Hugging Face lets us follow the models that are published from other models. A quantised version, an adaptation for one task or a model tuned on new data can keep a reference to the original. Following those references, we found 470,289 derivative repositories in our selection. Nearly half, 229,596, start from Qwen.

Figure 1.4

Derivative repositories and publishing accounts by model family

Derivative repositories and distinct publishing accounts, by base-model familyQwen: 229,596 derivative repositories (34% from accounts with 1,000 or more), 35,098 distinct accounts; Google: 72,718 derivative repositories (44% from accounts with 1,000 or more), 12,457 distinct accounts; Meta (Llama): 65,987 derivative repositories (12% from accounts with 1,000 or more), 15,496 distinct accounts; Mistral AI: 32,857 derivative repositories (19% from accounts with 1,000 or more), 8,285 distinct accounts; Hugging Face (Smol): 16,111 derivative repositories (3% from accounts with 1,000 or more), 4,436 distinct accounts; DeepSeek: 12,269 derivative repositories (17% from accounts with 1,000 or more), 4,349 distinct accounts; Microsoft: 10,762 derivative repositories (13% from accounts with 1,000 or more), 3,428 distinct accounts; Ai2: 3,558 derivative repositories (15% from accounts with 1,000 or more), 486 distinct accounts.Repositories · publishing accountsQwen229,596 · 35,098 accountsGoogle72,718 · 12,457 accountsMeta (Llama)65,987 · 15,496 accountsMistral AI32,857 · 8,285 accountsHugging Face (Smol)16,111 · 4,436 accountsDeepSeek12,269 · 4,349 accountsMicrosoft10,762 · 3,428 accountsAi23,558 · 486 accounts
Derivative repositories and distinct publishing accounts, by base-model familyQwen: 229,596 derivative repositories (34% from accounts with 1,000 or more), 35,098 distinct accounts; Google: 72,718 derivative repositories (44% from accounts with 1,000 or more), 12,457 distinct accounts; Meta (Llama): 65,987 derivative repositories (12% from accounts with 1,000 or more), 15,496 distinct accounts; Mistral AI: 32,857 derivative repositories (19% from accounts with 1,000 or more), 8,285 distinct accounts; Hugging Face (Smol): 16,111 derivative repositories (3% from accounts with 1,000 or more), 4,436 distinct accounts; DeepSeek: 12,269 derivative repositories (17% from accounts with 1,000 or more), 4,349 distinct accounts; Microsoft: 10,762 derivative repositories (13% from accounts with 1,000 or more), 3,428 distinct accounts; Ai2: 3,558 derivative repositories (15% from accounts with 1,000 or more), 486 distinct accounts.Repositories · publishing accountsQwen230k · 35k accountsGoogle73k · 12k accountsMeta (Llama)66k · 15k accountsMistral AI33k · 8.3k accountsHF (Smol)16k · 4.4k accountsDeepSeek12k · 4.3k accountsMicrosoft11k · 3.4k accountsAi23.6k · 0.5k accounts
  • China-based base model
  • US-based
  • Europe-based
  • from accounts with 1,000+ derivatives

Source Hugging Face repository metadata (hub-stats), 22 October 2026

Counted by the accounts that publish them, Qwen still leads: 35,098 accounts built on Qwen, more than on Meta’s Llama (15,496) and Google’s models (12,457) together. The shift is recent: in 2024, 8,446 accounts published derivatives of US-based models and 2,040 of China-based ones; in 2026, to 22 October, it was 13,352 against 25,574. An account is not a person, and some publish automatically.

How we measured it
A derivative is a public repository that declares a tracked model as its base, directly or through other repositories. Family = the base model’s developer namespace; accounts are distinct publishers, automated uploaders included. Deleted repositories are not seen, so early years are undercounted. *2026 to 22 October.
Source
Hugging Face repository metadata (base_model declarations) via the cfahlgren1/hub-stats archive, snapshot of 2026-10-22 (archive licensed Apache-2.0); region labels: ATOM
Base
470,289 derivative repositories alive on 2026-10-22
Evidence
E3our tabulation
Figure 1.4 data: derivative repositories and accounts
Family / yearBase regionRepositoriesFrom accounts with 1,000+ShareDistinct accountsLargest accountIts repositories
QwenChina-based229,59677,03033.6%35,098mradermacher12,205
googleUS-based72,71831,77243.7%12,457Krabat5,049
meta-llamaUS-based65,9878,04112.2%15,496mradermacher4,663
mistralaiEurope-based32,8576,34719.3%8,285mradermacher4,064
HuggingFaceTBUS-based16,1114542.8%4,436qualiaadmin746
deepseek-aiChina-based12,2692,11417.2%4,349mradermacher972
microsoftUS-based10,7621,38112.8%3,428GMorgulis578
allenaiUS-based3,55854415.3%486mradermacher337
openaiUS-based3,45870420.4%1,405dementor-research409
nvidiaUS-based3,10580225.8%936mradermacher378
2024 · all familiesChina-based80,92858,5752,040
2024 · all familiesEurope-based12,3581,8183,433
2024 · all familiesUS-based64,76530,0888,446
2025 · all familiesChina-based65,43810,12213,615
2025 · all familiesEurope-based12,7663,0462,704
2025 · all familiesUS-based60,8236,73214,422
2026 · all familiesChina-based103,25211,80725,574
2026 · all familiesEurope-based6,3161,4572,044
2026 · all familiesUS-based55,5777,90513,352

06 Other leaders

Beyond the downloads, DeepSeek, Z.ai and Moonshot stand out for other reasons

DeepSeek was the open lab with the most tokens on OpenRouter in 19 of the 21 complete months since January 2025, and took 31.3% of the open-weight tokens from October 2025 to September 2026. Z.ai, with GLM, took 12.1%, and GLM-5.3 Flash was the most used version in the top 50 in September 2026. Moonshot barely registers in that traffic, 3.7%, but its Kimi K3.1 Pro is the open model closest to the closed ones on the tests of chapter 02.

Who stands out on each dimension

  1. Downloads

    Qwen

    54.2% from Oct. 2025 to Sept. 2026; 55.4% in the October 2026 reading

    Hugging Face, the models we track

  2. Derivatives

    Qwen

    229,596 of 470,289 derivative repositories

    Hugging Face, repositories that declare a base model

  3. Tokens

    DeepSeek

    31.3% from Oct. 2025 to Sept. 2026; first in 19 of 21 months

    OpenRouter, open weights in the daily top 50

  4. Requests

    Meta (Llama)

    29.1% in October 2026

    Cloudflare Workers AI, text generation

  5. Capability

    Moonshot (Kimi K3.1 Pro)

    2.2 points from the best closed model in graduate-level science and 1.5 in long agentic tasks

    Epoch AI and Vals AI, best open model

  6. Cost per solved task

    DeepSeek and Z.ai (Flash versions)

    DeepSeek V4.1 Flash, $1.70 with 31.8% of tasks; GLM-5.3 Flash, $3.26 with 37.4%

    Vals AI, Terminal-Bench 4.0

  7. Research

    GPT and ChatGPT; then Llama

    65% and 23% in 2025; DeepSeek, from 0% to 14% since 2023

    OpenAlex, papers that name a family

Sources: Hugging Face hub-stats and the ATOM Project (chapter 01); OpenRouter (chapters 01 and 03); Cloudflare Radar, with permission (chapter 03); Epoch AI and Vals AI (chapter 02); Vals AI (chapter 04); OpenAlex (chapter 01).

The three profiles differ. DeepSeek combines volume and continuity: its versions change, but the lab stays in front. Z.ai competes with cheap versions that solve a reasonable share of long tasks. Kimi K3.1 Pro comes close to the closed models in graduate-level science and agentic tasks, but it still carries little of the use, and its cost is high:$7.18 per solved task on Terminal-Bench 4.0, four times as much as DeepSeek V4.1 Flash ($1.70).

07 What papers name

Which models appear in the scientific literature

Research papers tell another story. GPT and ChatGPT are still the most named, but their share of the papers that name a model family has fallen: 65% in 2025, down from 93% in 2023. Llama rose from 10% to 23%.

Share of the research papers that name a model family, by family, 2023 and 2025

2023 2025

0%25%50%75%100%GPT / ChatGPT65% from 93%Llama23% from 10%Gemini18% from 1%DeepSeek14% from 0%Claude12% from 2%Qwen6% from 0%
0%25%50%75%100%GPT / ChatGPT65% from 93%Llama23% from 10%Gemini18% from 1%DeepSeek14% from 0%Claude12% from 2%Qwen6% from 0%
Share of the research papers that name a model family, by family
Model family20232025
GPT / ChatGPT93%65%
Llama10%23%
Gemini1%18%
DeepSeek0%14%
Claude2%12%
Qwen0%6%
Papers in OpenAlex whose title or abstract names a model family together with a language-model term. One paper can name several families, so the shares add up to more than 100%. A mention is not use. Source: OpenAlex (CC0), queried 26 October 2026.

Qwen is rising in the literature too, from 0.3% of the papers that name a family in 2023 to 6% in 2025, but that is far from its 54% of the year’s downloads. DeepSeek was named in 14%, twice as often. Papers show what researchers study; downloads, what people run.

Chapter 01 · Closing

A cross-section of the downloads

Qwen has become the dominant name in the downloads, and much of that presence is concentrated in small models, which are easier to run on home hardware. But home use does not mean business use. The tokens served through APIs, in Chapter 03, tell a quite different story.

Qwen’s small models
Qwen has become the dominant name in the downloads, and much of that presence is concentrated in small models, which are easier to run on home hardware.
Downloaded and used
Each measure crowns a different lab: Qwen leads the downloads (54%), DeepSeek the tokens (31%) and Meta the requests on Workers AI (29%).
Home use, business use
Home use does not mean business use. The tokens served through APIs, in Chapter 03, tell a quite different story.