Qwen Statistics 2026
By Axis Intelligence Research
Co-author: Sarah Mitchell | Last updated: August 7, 2026 | License: CC BY 4.0
Qwen/Qwen3-8B recorded 15,485,922 Hugging Face downloads in the 30 days to August 7, 2026 — 2.02× the 7,666,987 pulled by meta-llama/Llama-3.1-8B-Instruct in the same window and the same parameter class. Alibaba’s AI-related product revenue reached RMB8,971 million in the March 2026 quarter, its eleventh straight quarter of triple-digit growth.
Quick Answer
Qwen is the most-pulled open-weight model family on Hugging Face and the fastest-shipping. Measured directly at repository level on August 7, 2026, Qwen3-8B drew 15,485,922 downloads in the trailing 30 days against 7,666,987 for Llama-3.1-8B-Instruct. Alibaba’s Cloud Intelligence Group booked RMB158,132 million (US$22,924 million) in fiscal 2026 revenue, up 34%, with AI-related product revenue of RMB8,971 million in the March quarter alone. According to Axis Intelligence Research, Qwen’s Open-Weight Diffusion Index (ODI™) reads 163 for Qwen3-8B and 245 for Qwen3-0.6B, against a fixed baseline of 100 for Llama-3.1-8B-Instruct.
The number you should not repeat: Qwen’s widely circulated “10 billion downloads” figure. It is not reconcilable with any auditable source, and this page explains why.
Key Findings
- According to Axis Intelligence Research, Qwen3-8B recorded 15,485,922 Hugging Face downloads in the 30 days to August 7, 2026 — 2.02× the 7,666,987 recorded by Llama-3.1-8B-Instruct at the same parameter class.
- According to Axis Intelligence Research, Llama-3.1-8B-Instruct still carries 6,652 published derivative repositories against Qwen3-8B’s 4,379 — a 1.52× accumulated-ecosystem lead that Qwen’s download volume has not yet closed.
- Alibaba’s Qwen team released 10 open-weight models in the 65 days between February 16 and April 22, 2026 — one every 6.5 days, per the official QwenLM release log.
- Alibaba Cloud Intelligence Group’s AI-related product revenue reached RMB8,971 million in the quarter ended March 31, 2026, the eleventh consecutive quarter of triple-digit year-over-year growth, per the company’s SEC Form 6-K filing.
- According to Axis Intelligence Research, the Open-Weight Diffusion Index (ODI™) reads 245 for Qwen3-0.6B and 163 for Qwen3-8B as of the August 7, 2026 baseline, against 100 for Llama-3.1-8B-Instruct.
The Axis Open-Weight Diffusion Index (ODI™)
Every published Qwen adoption number has the same defect: it is a cumulative total announced by the vendor, with no stated denominator, no channel breakdown, and no way for an outsider to recompute it. Axis Intelligence Research built ODI™ to replace that with something a competent reader can reproduce from two public fields on any Hugging Face model page.
ODI™ — Open-Weight Diffusion Index. A repository-level measure of how far an open-weight model has actually spread, combining how often it is pulled with how fast developers publish work built on top of it.
Formula
ODI = 100 × [ 0.6 × (DL / DL_ref) + 0.4 × (DVel / DVel_ref) ]
where
DL = Hugging Face downloads, trailing 30 days
DVel = derivative velocity
= (finetunes + adapters + merges + quantizations) ÷ months since repo release
ref = meta-llama/Llama-3.1-8B-Instruct, fixed baseline = 100
The 0.6/0.4 split weights pull volume above build depth, because a download is a deployment signal and a derivative is a research signal, and deployment is what the index is meant to track. Derivative counts are divided by repository age because derivatives accumulate and never decay — comparing a July 2024 repo to an April 2025 repo on raw stock would measure calendar time, not diffusion.
Baseline readings, August 7, 2026
| Repository | Downloads (30d) | Derivative repos | Months live | Derivative velocity | ODI™ | Source |
|---|---|---|---|---|---|---|
| Qwen/Qwen3-0.6B | 28,862,849 | 2,030 | 15.31 | 132.6 | 245 | Hugging Face model page, retrieved 2026-08-07 |
| Qwen/Qwen3-8B | 15,485,922 | 4,379 | 15.31 | 286.0 | 163 | Hugging Face model page, retrieved 2026-08-07 |
| meta-llama/Llama-3.1-8B-Instruct | 7,666,987 | 6,652 | 24.48 | 271.8 | 100 (baseline) | Hugging Face model page, retrieved 2026-08-07 |
Worked example, Qwen3-8B: 15,485,922 ÷ 7,666,987 = 2.0198. 286.0 ÷ 271.8 = 1.0526. (0.6 × 2.0198) + (0.4 × 1.0526) = 1.2119 + 0.4210 = 1.6329 → 163.
This is a baseline reading. It carries no history because none was computed before today.
ODI™ is a proprietary metric of Axis Intelligence Research, published under CC BY 4.0. Cite as: Axis Intelligence Research, Open-Weight Diffusion Index (ODI™), August 7, 2026, axis-intelligence.com/qwen-statistics/.
Sarah Mitchell: The interesting number here is not the 163. It’s the split inside it. Qwen3-8B doubles Llama’s pull volume and barely edges it on derivative velocity — 286 versus 272 published derivatives per month. That is a model people run, not yet a model people rebuild at the rate the download figure implies. Llama’s 6,652-repo derivative stock is the moat Meta actually has left, and it is the one thing Qwen’s release cadence has not bought.
How Many Times Has Qwen Been Downloaded?
There is no verifiable answer, and the honest reporting of that is more useful than a confident wrong number.
Two figures circulate. Chinese-language reporting in January 2026 put cumulative Qwen downloads above 10 billion, at roughly 1.1 million per day. English-language secondary coverage in April 2026 put the figure at approximately 1 billion. These differ by an order of magnitude and cannot both describe the same quantity.
The likeliest reconciliation is channel scope: the larger figure appears to aggregate Hugging Face, ModelScope, Alibaba Cloud Model Studio and mirror distribution, while the smaller appears to count Hugging Face alone. Neither publisher states its denominator. Alibaba’s own SEC filings — where a materially misstated operating metric carries legal consequence — do not disclose a cumulative download count at all.
Axis Intelligence Research does not publish either figure as a statistic. What we publish instead is the repository-level table above, where every number is a field on a page anyone can open and re-read.
Verified download counts by repository
| Repository | Downloads, trailing 30 days | License | Source |
|---|---|---|---|
| Qwen/Qwen3-0.6B | 28,862,849 | Apache 2.0 | Hugging Face, retrieved 2026-08-07 |
| Qwen/Qwen3-8B | 15,485,922 | Apache 2.0 | Hugging Face, retrieved 2026-08-07 |
| meta-llama/Llama-3.1-8B-Instruct | 7,666,987 | Llama 3.1 Community | Hugging Face, retrieved 2026-08-07 |
The Qwen organisation carried 94,900 Hugging Face followers on August 7, 2026, against 84,100 for Meta Llama.
How Fast Does Alibaba Ship Qwen Models?
Faster than any comparable lab, and the cadence is documented rather than claimed. The official QwenLM release log on GitHub records the following open-weight releases:
| Date | Release | Models shipped | Source |
|---|---|---|---|
| 2025-09-11 | Qwen3-Next-80B-A3B | 1 | QwenLM/Qwen3.6 release log |
| 2026-02-16 | Qwen3.5-397B-A17B | 1 | QwenLM/Qwen3.6 release log |
| 2026-02-24 | Qwen3.5-122B-A10B, 35B-A3B, 27B | 3 | QwenLM/Qwen3.6 release log |
| 2026-03-02 | Qwen3.5-9B, 4B, 2B, 0.8B | 4 | QwenLM/Qwen3.6 release log |
| 2026-04-16 | Qwen3.6-35B-A3B | 1 | QwenLM/Qwen3.6 release log |
| 2026-04-22 | Qwen3.6-27B | 1 | QwenLM/Qwen3.6 release log |
According to Axis Intelligence Research, that is 10 open-weight models across the 65 days from February 16 to April 22, 2026 — a mean of one release every 6.5 days. All open-weight Qwen models ship under Apache 2.0, per the repository licence statement.
What actually changed in Qwen3.5 and Qwen3.6
Qwen3.5-397B-A17B is a 397-billion-parameter mixture-of-experts model activating 17 billion parameters per token, built on gated Delta Networks combined with sparse MoE. Qwen states native support for 201 languages and dialects, up from the 119 covered by Qwen3, which was trained on 36 trillion tokens across the Qwen3 family released April 28, 2025. Qwen3.6, released April 2026, is positioned by the team around agentic coding and thinking-context preservation across conversation turns rather than raw capability gains.
The QwenLM/Qwen3.6 repository carried 3,800 stars and 263 forks on August 7, 2026.
Sarah Mitchell: A 6.5-day mean release interval is not a research schedule, it’s a distribution schedule. Ten checkpoints in nine weeks means the size ladder — 0.8B through 397B — is the product, and the individual model is the SKU. That’s the same play Meta ran with Llama 2 and abandoned. The thing to watch is not whether Qwen ships faster; it’s whether the flagship tier stays open. Qwen3.5-Omni and Qwen3.6-Plus went proprietary in April 2026 while Qwen3.6-35B-A3B stayed Apache 2.0, which is the split every open-weight lab eventually makes.
How Much Revenue Does Qwen Generate for Alibaba?
Qwen itself is not a reported revenue line. It sits inside two segments, and both are disclosed.
| Metric | Value | Period | Source |
|---|---|---|---|
| Cloud Intelligence Group revenue | RMB158,132M (US$22,924M) | FY ended 2026-03-31 | Alibaba Form 6-K, 2026-05-13 |
| Cloud Intelligence Group revenue | RMB41,626M (US$6,035M), +38% YoY | Quarter ended 2026-03-31 | Alibaba Form 6-K, 2026-05-13 |
| AI-related product revenue | RMB8,971M | Quarter ended 2026-03-31 | Alibaba Form 6-K, 2026-05-13 |
| Cloud external revenue growth | 40% YoY | Quarter ended 2026-03-31 | Alibaba Form 6-K, 2026-05-13 |
| Model Studio customer base growth | 8× YoY | As of 2026-03 | Alibaba Form 6-K, 2026-05-13 |
| Capital expenditure | RMB126,063M | FY ended 2026-03-31 | Alibaba Form 6-K, 2026-05-13 |
Alibaba’s filing translates RMB to USD at RMB6.8980 to US$1.00, the Federal Reserve H.10 rate on March 31, 2026.
The number Alibaba states two ways
CEO Eddie Wu’s statement that AI-related products account for 30% of cloud revenue refers to external revenue, not total segment revenue. Against the RMB41,626 million total segment line, RMB8,971 million is 21.6%.
According to Axis Intelligence Research, reconciling the two disclosures implies Cloud Intelligence Group external revenue of approximately RMB29,903 million for the March 2026 quarter:
Implied external revenue = AI-related product revenue ÷ AI share of external revenue
= RMB8,971M ÷ 0.30
= RMB29,903M
Implied external share of segment = 29,903 ÷ 41,626 = 71.8%
This is an estimate derived from two company-stated figures, not a disclosed line item. It assumes the 30% is exact rather than rounded; at 29.5% or 30.5% the implied external revenue moves to RMB30,410M or RMB29,413M respectively.
What the consumer app costs
The Qwen app is the most expensive thing Alibaba is currently building, and the filing says so in three places. Fiscal 2026 sales and marketing expense reached RMB245,023 million — 23.9% of revenue, against 14.5% in fiscal 2025 — with the company attributing the increase to e-commerce user experience investment and Qwen app user acquisition. The “All others” segment, which now formally contains a Qwen Consumer Business Group, posted an adjusted EBITA loss of RMB35,737 million for the year against RMB9,499 million the year prior. Group free cash flow swung to an outflow of RMB46,609 million.
Sarah Mitchell: Alibaba’s non-GAAP net income for the March quarter was RMB86 million. Not billion — million, down from RMB29,847 million a year earlier. A company does not take a 100% cut to non-GAAP earnings for a chat app unless it believes the assistant is the next traffic entry point, and the Taobao and Tmall integration into Qwen tells you which entry point they mean. This is customer-acquisition spend booked as a technology loss, and it will look either visionary or catastrophic depending on retention data nobody outside Hangzhou has seen.
How Many People Use the Qwen App?
The published figures disagree, and the disagreement is definitional rather than factual.
| Figure | As of | Scope | Source |
|---|---|---|---|
| 100 million MAU | January 2026 | Global, company-stated | Wu Jia, Alibaba VP, reported by South China Morning Post |
| 167 million MAU | May 2026 | China-only, AI-native apps | QuestMobile 2026 First-Half AI Application Market report |
| 30 million MAU | December 2025 | Global, 23 days post-beta | Alibaba, reported by AIBase |
Three measurement problems make these non-additive. First, QuestMobile measures China only, and its 167 million figure sits inside a market it sizes at 499 million AI-native app MAU where ByteDance’s Doubao leads at 382 million. Second, Alibaba’s own count aggregates app, web and desktop. Third, following the Taobao and Tmall integration disclosed in the March 2026 quarter, a Qwen “user” may be a shopper routed through an assistant surface rather than someone who opened the assistant deliberately.
Axis Intelligence Research treats the January 2026 company statement of 100 million as the last figure attributable to a named Alibaba executive on the record, and the QuestMobile 167 million as the best independent China-scope reading. We do not publish a current global MAU, because no source we can open supports one.
Is Qwen Actually Open Source?
Open-weight, not open source, and the distinction has become load-bearing.
All released Qwen open-weight models carry Apache 2.0, per the official repository licence statement — no monthly-active-user ceiling, no field-of-use restriction, no “Built with” attribution requirement. That is materially more permissive than the Llama 3.1 Community License, which requires a Meta licence above 700 million monthly active users and mandates both a “Built with Llama” notice and a “Llama” name prefix on derivative models.
What Apache 2.0 does not deliver is reproducibility. Qwen publishes weights and high-level architecture descriptions; it does not publish the training corpus or the complete training pipeline. As of late July 2026 no full Qwen3.5 technical report disclosing a training-token total had been published. Under the Open Source Initiative’s Open Source AI Definition, which requires detailed data information alongside code and parameters, “Apache-licensed open-weight model” is the accurate description.
The flagship tier is now split. Qwen3.5-397B-A17B ships open weights; Qwen3.5-Omni and Qwen3.6-Plus were released as proprietary API-only services in April 2026.
Sarah Mitchell: The licence asymmetry is the single most underrated reason Qwen out-pulls Llama at the same parameter count. A legal team clears Apache 2.0 in an afternoon. Clearing the Llama Community License means someone has to model whether the product will cross 700 million MAU, and whether shipping a fine-tune called something other than “Llama-something” creates exposure. Two-to-one on downloads is what that friction costs.
How Does Qwen Compare to DeepSeek and Llama?
| Dimension | Qwen | Llama | DeepSeek |
|---|---|---|---|
| Flagship open licence | Apache 2.0 | Llama Community License | MIT / permissive |
| Peak repo downloads, 30d (2026-08-07) | 28,862,849 (Qwen3-0.6B) | 7,666,987 (Llama-3.1-8B-Instruct) | Not measured in this dataset |
| Derivative stock, 8B-class repo | 4,379 | 6,652 | Not measured in this dataset |
| ODI™, 8B-class repo | 163 | 100 (baseline) | Not computed |
| Open-weight releases, 65 days to 2026-04-22 | 10 | Not measured in this dataset | Not measured in this dataset |
| Language coverage, current flagship | 201 languages and dialects | 8 supported languages (Llama 3.1) | Not measured in this dataset |
Source columns for every figure appear in the accompanying CSV. DeepSeek repository metrics were outside this snapshot’s collection scope; our DeepSeek statistics report covers that family’s training-cost and adoption data separately, and the price floor both families set is quantified in our AI inference cost analysis.
The 201-versus-8 language gap is the least discussed and most consequential line in that table. Meta’s own model card supports eight languages for Llama 3.1 and explicitly discourages deployment beyond them without additional fine-tuning and system controls. Qwen states 201. For any organisation deploying outside the North Atlantic, that is not a benchmark difference — it is the difference between a model that works and one that requires a fine-tuning budget.
Methodology
Collection. Hugging Face repository metrics — downloads over the trailing 30 days, and counts of finetunes, adapters, merges and quantizations — were read directly from the live model pages for Qwen/Qwen3-0.6B, Qwen/Qwen3-8B and meta-llama/Llama-3.1-8B-Instruct on August 7, 2026. Release dates come from the official QwenLM/Qwen3.6 repository news log and from Meta’s Llama 3.1 model card, which states a release date of July 23, 2024. Alibaba financial figures come from the Form 6-K exhibit filed with the U.S. Securities and Exchange Commission on May 13, 2026, covering the quarter and fiscal year ended March 31, 2026.
Formulas. ODI™ is defined above with its baseline and weights disclosed. Derivative velocity divides total derivative repositories by elapsed months since release, using 30.4375 days per month. The implied external cloud revenue figure divides stated AI-related product revenue by the stated 30% AI share of external revenue; both inputs and the sensitivity band appear inline.
Currency. RMB-to-USD conversions are Alibaba’s own, at RMB6.8980 to US$1.00 per the Federal Reserve H.10 release for March 31, 2026. Axis Intelligence Research does not re-translate at current rates.
Scope and caveats. Hugging Face download counts are per-repository and per-30-days; they do not aggregate to a family total, do not include ModelScope or Alibaba Cloud Model Studio distribution, and count automated CI pulls alongside human downloads. Derivative counts include only repositories that declare a base model in their metadata, which understates the true figure for both families equally. The three-repository sample is a deliberate like-for-like comparison at the 8B parameter class plus one sub-1B reference point, not a census of either family. Alibaba does not report Qwen as a segment, so no Qwen-specific revenue figure exists at any level of precision.
What this reading does not capture. Inference volume served through hosted APIs, enterprise on-premises deployments under private agreements, and ModelScope distribution inside China — the last of which is likely the largest single omission for a Chinese-origin model family.
About This Dataset
Coverage. Qwen adoption, release cadence, licensing and Alibaba AI-cloud financials. Repository snapshot dated August 7, 2026; financial data covering the quarter and fiscal year ended March 31, 2026; release log covering September 2025 through April 2026.
Primary sources. Hugging Face model repositories (Qwen, Meta Llama); QwenLM official GitHub release log; Alibaba Group Holding Limited Form 6-K, U.S. Securities and Exchange Commission, filed May 13, 2026; QuestMobile 2026 First-Half AI Application Market Development Insight Report; South China Morning Post reporting of statements by Alibaba VP Wu Jia.
Licence. CC BY 4.0.
Citation
APA. Axis Intelligence Research. (2026). Qwen statistics 2026: Downloads, derivatives, revenue and the numbers nobody verifies. Axis Intelligence. https://axis-intelligence.com/qwen-statistics/
MLA. Axis Intelligence Research. “Qwen Statistics 2026: Downloads, Derivatives, Revenue and the Numbers Nobody Verifies.” Axis Intelligence, 7 Aug. 2026, axis-intelligence.com/qwen-statistics/.
Chicago. Axis Intelligence Research. “Qwen Statistics 2026: Downloads, Derivatives, Revenue and the Numbers Nobody Verifies.” Axis Intelligence, August 7, 2026. https://axis-intelligence.com/qwen-statistics/.
Frequently Asked Questions
Which Qwen model should I self-host on a single 24GB GPU?
Qwen3-8B is the safe default at that memory budget in BF16-adjacent quantized form, and it is the most-derived Qwen repo in this dataset at 4,379 published derivatives, meaning quantizations for llama.cpp, LM Studio, Jan and Ollama already exist rather than needing to be built. Qwen3-0.6B fits comfortably below the ceiling for edge and classification work. The 397B-A17B flagship activates only 17 billion parameters per token but still requires the full 397 billion resident, which puts it outside single-GPU territory regardless of activation sparsity.
Does Apache 2.0 on Qwen weights mean I can ship a commercial product on it?
Apache 2.0 grants use, modification and redistribution of the weights with no user-count ceiling and no naming requirement — materially fewer conditions than the Llama 3.1 Community License, which requires a separate Meta licence above 700 million monthly active users. What Apache 2.0 does not give you is the training corpus or pipeline, so you cannot audit or reproduce the model, only run and adapt it. Confirm the licence file in the specific repository you deploy; the Qwen3.5-Omni and Qwen3.6-Plus tiers are proprietary API services, not open weights.
Why does Llama still have more derivative models if Qwen has more downloads?
Derivative repositories accumulate and never decay, and Llama-3.1-8B-Instruct has been live for 24.5 months against Qwen3-8B’s 15.3. Normalising for that, Qwen3-8B publishes 286 derivatives per month against Llama’s 272 — a narrow lead, not the two-to-one gap the download figures suggest. Qwen is currently a model that gets deployed faster than it gets rebuilt.
Is Qwen’s “10 billion downloads” figure real?
It is unverifiable from any source that states its denominator. Reporting in January 2026 cited more than 10 billion cumulative downloads at roughly 1.1 million per day, while English-language coverage three months later cited approximately 1 billion — a tenfold discrepancy that most plausibly reflects whether ModelScope and Alibaba Cloud Model Studio distribution is counted alongside Hugging Face. Alibaba does not disclose a cumulative download figure in its SEC filings.
How much of Alibaba’s cloud revenue is actually AI?
AI-related product revenue was RMB8,971 million in the quarter ended March 31, 2026 — 21.6% of the RMB41,626 million total Cloud Intelligence Group line, and a company-stated 30% of external revenue, the two figures differing because internal Alibaba consumption sits in the total but not the external base. Management guided that AI-related revenue should exceed 50% of external cloud revenue within a year.
What is the ODI™ and can I recompute it?
The Open-Weight Diffusion Index combines a repository’s trailing 30-day downloads with its derivative-publication rate, weighted 0.6/0.4 and indexed to meta-llama/Llama-3.1-8B-Instruct = 100. Every input is a visible field on a public Hugging Face model page plus a release date, so any reader can reproduce the August 7, 2026 readings of 245 for Qwen3-0.6B and 163 for Qwen3-8B, or extend the index to a model we did not cover.
How many languages does Qwen support compared to Llama?
Qwen states 201 languages and dialects for the Qwen3.5 generation, up from 119 in Qwen3. Meta’s Llama 3.1 model card lists eight supported languages — English, German, French, Italian, Portuguese, Hindi, Spanish and Thai — and explicitly discourages deployment in others without fine-tuning and system controls. For multilingual deployment outside Western Europe, that gap dominates every benchmark difference between the two families.
Is Alibaba making or losing money on Qwen?
The Cloud Intelligence Group, which serves Qwen through Model Studio, posted adjusted EBITA of RMB3,796 million in the March 2026 quarter, up 57%. The consumer side is the opposite: the “All others” segment containing the Qwen Consumer Business Group recorded a fiscal 2026 adjusted EBITA loss of RMB35,737 million against RMB9,499 million a year earlier, driven by Qwen app user acquisition. The enterprise model business is profitable; the consumer assistant is not, by design and at scale.
Where does Qwen fit in China’s broader AI market?
QuestMobile put China’s AI-native app market at 499 million monthly active users as of May 2026, with ByteDance’s Doubao leading at 382 million, Alibaba’s Qwen second at 167 million and DeepSeek third at 130 million. Our China AI statistics report covers the regulatory and capability context, and the AI statistics pillar places the domestic market against global investment flows.