Unboxing the AI Stack
L02 · Language Models · Feature

Chinese Open-Weight Models: Downloads, Derivatives and Token Share

Chinese models passed US models in cumulative Hugging Face downloads in summer 2025, and Qwen is now the most common base for fine-tunes. Downloads, derivatives, token share and licenses, each study's figure and method.

Hugging Face downloads, last 30 days millions

Qwen3-0.6B30.73
GLM-5.3-Flash6.51
DeepSeek-V3.24.78
Kimi-K31.12
MiniMax-M2.70.70
Hy30.58

Measured: Hugging Face's own counts, retrieved Oct 13, 2026

Chart: downloads over the last 30 days of the most-downloaded general language model repo from each of six Chinese labs' official Hugging Face accounts, as counted by Hugging Face on Oct 13, 2026. Full table below.

As of Download and like counts measured by Hugging FaceLicenses as stated on each model card

Most-downloaded general language model repo per Chinese lab on Hugging Face, as of 2026-10-13
Lab Family Repo Repo created License Downloads (last 30 days) Likes Figure
Alibaba Qwen8 Qwen/Qwen3-0.6B 2025-04-27 Apache 2.0 30,726,361 1,805 Measured
Zhipu AI (Z.ai) GLM12 zai-org/GLM-5.3-Flash 2026-08-25 MIT 6,511,401 2,832 Measured
DeepSeek DeepSeek9 deepseek-ai/DeepSeek-V3.2 2025-12-01 MIT 4,779,548 1,571 Measured
Moonshot AI Kimi11 moonshotai/Kimi-K3 2026-06-13 Kimi K3 License 1,117,095 11,707 Measured
MiniMax MiniMax14 MiniMaxAI/MiniMax-M2.7 2026-04-09 Non-commercial license 702,817 1,256 Measured
Tencent Hy18 tencent/Hy3 2026-07-02 Apache 2.0 578,301 979 Measured
StepFun Step17 stepfun-ai/Step-3.5-Flash 2026-02-01 Apache 2.0 264,837 840 Measured
Xiaomi MiMo16 XiaomiMiMo/MiMo-V2.5 2026-04-27 MIT 192,784 431 Measured
Baidu ERNIE19 baidu/ERNIE-4.5-21B-A3B-Thinking 2025-09-08 Apache 2.0 89,211 793 Measured

One row per lab: the most-downloaded general-purpose language model (text or vision-language chat model) in the lab's official Hugging Face organization, by Hugging Face's "downloads last month" count, pulled from the Hub API on Oct 13, 2026.7 OCR, embedding, translation, image and video models and quantized copies are left out, so DeepSeek-OCR, Qwen3-Embedding, MiniMax-H3 (video) and Tencent's Hy-MT2 translation GGUFs, each of which out-downloads some rows here, don't appear. Downloads count every request for a repo's query file, CI pipelines and automated tests included, and aren't unique users;6 they favor small models (see "What a download measures"), so this table is a dated snapshot, not a ranking of the labs. "Repo created" is the Hub's creation date, which can precede the public release. Licenses are named as on each model card; "Non-commercial license" is MiniMax's own heading for M2.7's terms. Figure labels: how they work.

Chinese open-weight models, by the numbers

Downloads: three studies, three shares

An open-weight model is one whose trained weights anyone can download and run, whatever the license says about using them. Hugging Face is where most of those downloads happen, and three recent studies have counted them by country of the model's developer. All three find Chinese models ahead of US models, but each measures something different, so their shares can't be compared with each other.

The ATOM Report by Nathan Lambert and Florian Brand tracks about 1,500 "mainline" open language models and their cumulative downloads from November 2023 to March 2026. By its count China overtook the US in late July 2025 and reached 1.15 billion cumulative downloads versus 723 million by March 2026 Measured, with all tracked models together reaching 2.04 billion, six times the 339 million of March 2025.2 Qwen alone accounts for 942.1 million of those downloads; it passed Meta's Llama in September 2025 (325.4 million versus 323.7 million) and by March 2026 had nearly double Llama's 476.0 million.2 DeepSeek passed Mistral in January 2026, at 128.2 million versus 124.3 million.2

Hugging Face's own "State of Open Source: Spring 2026" report, published March 17, 2026, says China passed the US in both monthly and overall downloads and that "in the past year, Chinese models quickly accounted for the plurality or 41% of downloads."1 Shayne Longpre and co-authors, working from the Hub's complete weekly download history from June 2020 to August 2025 (851,000 models), put China's share of downloads of models created between August 2024 and August 2025 at 17.1%, ahead of the US at 15.7% for the first time, with DeepSeek and Qwen alone taking 14%.3 One difference is scope: Longpre's study covers every model on the Hub, including those of quantizers and community re-uploaders such as lmstudio-community and mlx-community, which by its count took over 22% of recent downloads, while ATOM tracks a curated list of mainline models.32

Derivatives: Qwen as the default base model

A derivative is a model built on another one's weights: a fine-tune, an adapter, a merge. Developers pick a base model to fine-tune once, so derivatives show which weights people build on, not just which they try. On that measure, Alibaba has more derivative models on Hugging Face than Google and Meta combined, and the Qwen family alone has more than 113,000; counting every model that tags Qwen, the figure is over 200,000, according to Hugging Face.1

The ATOM Report, which counts derivatives through the Hub's base_model tag and excludes GGUF and MLX re-uploads, finds China's share of new derivatives rising from 10% in November 2023 to 70% by February 2026. Qwen's own share of new fine-tunes and adaptations went from 1% in January 2024 to 69% in February 2026, while Meta's peaked at 44% in August 2024 and fell to 11%.2

The growth isn't only Qwen. Hugging Face reports that Baidu went from zero releases on the Hub in 2024 to over 100 in 2025, that ByteDance and Tencent each increased releases eight- to ninefold, and that Baidu and MiniMax, which had favored closed approaches, shifted toward open releases.1 In its count of the most-liked models, DeepSeek-R1 went to the top within a year.1

Token share: what inference shows

Downloads show who fetched the weights; tokens processed show who is actually running a model. OpenRouter, a service that routes API requests to hundreds of models, published a study with a16z in December 2025 covering more than 100 trillion tokens.4 Over its one-year window to November 2025, Chinese open-source models averaged about 13.0% of weekly token volume across all models, open and closed, starting from a weekly low of 1.2% in late 2024 and reaching "nearly 30%" in some weeks. Open models from the rest of the world averaged 13.7%, and proprietary models outside China about 70% Measured.4 The study counts models from China, Taiwan and Hong Kong as Chinese.4

Within open models only, the shift is larger. Using monthly OpenRouter data on the top 10 open models, the ATOM Report finds Chinese models' share of open-model tokens rising from 2.8% to 72.7% by January 2026, with DeepSeek at 31.1% and Xiaomi's MiMo-V2-Flash, released in December 2025, at 27.2%.2

The current picture is on OpenRouter's live leaderboard. In the week of usage data through Oct 12, 2026, the five models with the most tokens processed were all from Chinese labs: DeepSeek V4.1 Flash (40.5 trillion tokens), GLM 5.3 Flash from Z.ai (12.7 trillion), StepFun's Step 5 Preview (12.1 trillion), Xiaomi's MiMo-V2.6-Flash (11.5 trillion) and Tencent's Hy4 preview (8.31 trillion) Measured.5 The leaderboard ranks models by usage, not openness, and doesn't say which have published weights. Two that have are DeepSeek-V4.1-Flash, released on Hugging Face under the MIT License with 1,272,612 downloads in the last 30 days, and GLM-5.3-Flash, also MIT.10712 OpenRouter notes that its rankings count both prompt and completion tokens, cover only traffic routed through it, and measure neither users nor spend.5

Licenses: how open is "open"

"Open-weight" says nothing about the terms. Among the repos in the table, Qwen3-0.6B, Tencent's Hy3, StepFun's Step 3.5 Flash and Baidu's ERNIE 4.5 21B-A3B Thinking are under Apache 2.0, and DeepSeek-V3.2, GLM-5.3-Flash and Xiaomi's MiMo-V2.5 under MIT, both permissive licenses that allow commercial use.891216171819

The largest flagships carry their own licenses. Moonshot AI's Kimi K3, which its card describes as a 2.8-trillion-parameter model and the first open "3T-class" model Vendor-stated, uses the Kimi K3 License: MIT-style permissions, but a Model-as-a-Service operator whose revenue, with its affiliates, exceeds $20 million over 12 months needs a separate agreement before commercial use, and products with more than 100 million monthly active users or $20 million in monthly revenue must display "Kimi K3".11 Z.ai's full-size GLM-5.3 uses a GLM-5.3 License with a similar structure, under which a Model-as-a-Service business above $10 billion in 12-month revenue must pass Z.ai's security review first, while its smaller GLM-5.3-Flash sibling is MIT.1312

MiniMax restricts commercial use more. MiniMax-M2.7 ships under a license headed "Non-Commercial License": personal, non-profit, academic and research use is free, and any commercial use needs MiniMax's prior written authorization.14 MiniMax-M3, described on its card as having about 428 billion parameters with about 23 billion activated Vendor-stated, uses the "MiniMax Community License": commercial users must display "Built with MiniMax M3", and products or services with more than $20 million in yearly revenue need MiniMax's prior written authorization, while smaller ones send a one-time notice.15 The download and token counts above include models under all of these terms alike.

What a download measures

Hugging Face counts a download as any HTTP request, GET or HEAD, to a repo's query file, by default config.json. It doesn't deduplicate users or exclude CI/CD pipelines; organizations that want that need its separate request logs.6 Hugging Face's own report says automated systems and CI pipelines inflate small-model counts, though it adds that the trend toward smaller models is real.1

That skews the numbers toward Qwen's small models. The ATOM Report finds that six small Qwen3 models, from 0.6B to 8B parameters, get as many monthly downloads as Zhipu AI, MiniMax, Mistral, Moonshot, NVIDIA and OpenAI combined, and that models under 10B parameters take about 75% of all downloads in its tracked set.2 In the table above, the most-downloaded Qwen repo is Qwen3-0.6B, one of those small models, with 30,726,361 downloads in 30 days, more than four times GLM-5.3-Flash's 6,511,401 in the next row.7 The ATOM authors say the large-model labs (MiniMax, Moonshot AI, Z.ai) "are far from dominating the adoption metrics of open models," and DeepSeek's influence is larger on inference platforms than in downloads.2

Token counts have their own limits: OpenRouter's data covers one platform, leaving out enterprise, local and internal deployments, and its authors call their results "indicative behavioral patterns rather than definitive measurements."4 Read together, the studies agree on the direction since mid-2025, and disagree on the size, because each counts a different slice: tracked mainline models, all of the Hub, or one inference router.

Sources

  1. Hugging Face, State of Open Source on Hugging Face: Spring 2026, published 2026-03-17, retrieved 2026-10-13.
  2. Nathan Lambert and Florian Brand, The ATOM Report: Measuring the Open Language Model Ecosystem, arXiv:2604.07190v2 (2026-05-25), retrieved 2026-10-13.
  3. Shayne Longpre et al., Economies of Open Intelligence: Tracing Power & Participation in the Model Ecosystem, arXiv:2512.03073v1 (2025-11-27), retrieved 2026-10-13.
  4. Malika Aubakirova et al. (a16z and OpenRouter), State of AI: An Empirical 100 Trillion Token Study with OpenRouter, December 2025, retrieved 2026-10-13.
  5. OpenRouter, AI Model Rankings (weekly view, usage data through 2026-10-12; CC BY 4.0), retrieved 2026-10-13.
  6. Hugging Face, Models Download Stats (Hub documentation), retrieved 2026-10-13.
  7. Hugging Face, Hub API, models endpoint, per-organization listings sorted by downloads for Qwen, deepseek-ai, moonshotai, zai-org, MiniMaxAI, tencent, stepfun-ai, XiaomiMiMo and baidu, retrieved 2026-10-13.
  8. Qwen (Alibaba), Qwen3-0.6B model card, retrieved 2026-10-13.
  9. DeepSeek, DeepSeek-V3.2 model card, retrieved 2026-10-13.
  10. DeepSeek, DeepSeek-V4.1-Flash model card, retrieved 2026-10-13.
  11. Moonshot AI, Kimi-K3 model card and Kimi K3 License, retrieved 2026-10-13.
  12. Z.ai, GLM-5.3-Flash model card, retrieved 2026-10-13.
  13. Z.ai, GLM-5.3 model card and GLM-5.3 License, retrieved 2026-10-13.
  14. MiniMax, MiniMax-M2.7 model card and license, retrieved 2026-10-13.
  15. MiniMax, MiniMax-M3 model card and MiniMax Community License, retrieved 2026-10-13.
  16. Xiaomi, MiMo-V2.5 model card, retrieved 2026-10-13.
  17. StepFun, Step 3.5 Flash model card, retrieved 2026-10-13.
  18. Tencent, Hy3 model card, retrieved 2026-10-13.
  19. Baidu, ERNIE-4.5-21B-A3B-Thinking model card, retrieved 2026-10-13.