Hugging Face counts 151,448 Qwen-based derivative models on its Hub. QuestMobile puts the Qwen app at 167 million monthly active users in China.
Those figures measure different things, and these Qwen AI statistics keep them apart. The post covers Qwen3.8-Max specs and benchmarks, API pricing, open-model adoption, app usage and Alibaba’s June-quarter 2026 results.
Qwen AI Statistics
2.4 Trillion Parameters: Qwen AI’s 2026 Flagship
Alibaba launched Qwen3.8-Max in August 2026, within three months of the prior version. The open weights, Qwen3.8-2.4T-A95B, went live on Hugging Face on August 12, 2026.
The mixture-of-experts model activates 95 billion of its 2.4 trillion parameters per token.
| Specification | Qwen3.8-2.4T-A95B |
|---|---|
| Total parameters | 2.4 trillion |
| Activated parameters | 95 billion |
| Experts | 512 (10 routed + 1 shared activated) |
| Layers | 92 |
| Hidden dimension | 8,192 |
| Context length | 262,144 native, extensible to 1,010,000 tokens |
The open checkpoint is text-only and cannot turn thinking off. The hosted Qwen3.8-Max adds vision input, a non-thinking mode, built-in tools and a 1M-token default context.
Qwen Open-Weight Releases in 2026
| Date | Release |
|---|---|
| Feb 16, 2026 | Qwen3.5, starting with a 397B-A17B MoE model |
| Feb 24, 2026 | Qwen3.5-122B-A10B, Qwen3.5-35B-A3B, Qwen3.5-27B |
| Mar 2, 2026 | Qwen3.5-9B, 4B, 2B and 0.8B |
| Apr 16, 2026 | Qwen3.6-35B-A3B |
| Apr 22, 2026 | Qwen3.6-27B |
| Aug 12, 2026 | Qwen3.8-2.4T-A95B |
How Does Qwen3.8-Max Score Against Claude?
Qwen3.8-Max beats Claude Opus 4.8 on four of the six benchmarks below, based on the Qwen team’s own results table. Claude Fable 5 leads Qwen3.8-Max on SWE-bench Pro, FrontierSWE and HLE.
| Benchmark | Qwen3.7-Max | Qwen3.8-Max | Claude Opus 4.8 | Claude Fable 5 |
|---|---|---|---|---|
| Terminal Bench 2.1 | 74.5 | 86.6 | 84.6 | 84.6 |
| SWE-bench Pro | 60.6 | 67.7 | 69.2 | 80.0 |
| FrontierSWE | 40.7 | 73.5 | 70.0 | 88.8 |
| PaperBench | 64.8 | 93.0 | 80.3 | 88.8 |
| GPQA Diamond | 92.4 | 92.6 | 92.0 | 92.6 |
| HLE | 41.4 | 43.6 | 45.7 | 53.3 |
Qwen3.8-Max leads Fable 5 on PaperBench, 93.0 to 88.8, and on Terminal Bench 2.1, 86.6 to 84.6. The two tie on GPQA Diamond at 92.6.
Between versions, FrontierSWE moved from 40.7 to 73.5 while GPQA Diamond moved from 92.4 to 92.6. The model card notes that Fable 5 results may involve fallbacks.
Opus 4.8 and Fable 5 are Anthropic models, and their traffic and revenue data sit in these Claude usage trends.
Earlier open code models were scored on older tests such as HumanEval, as these StarCoder benchmark figures show.
Qwen AI Pricing Statistics: $2 In, $6 Out
| Model | Input tokens per request | Input (per 1M) | Output (per 1M) |
|---|---|---|---|
| qwen3.8-max | 0–1M | 2 | 6 |
| qwen3.8-max-0902 | 0–1M | 2 | 6 |
| qwen3.7-max | 0–1M | 2.5 | 7.5 |
| qwen3.6-max-preview | 0–128K | 1.3 | 7.8 |
| qwen3-max | 0–32K | 1.2 | 6 |
| qwen3-max | 32K–128K | 2.4 | 12 |
| qwen3-max | 128K–256K | 3 | 15 |
| qwen3.7-plus | 0–256K | 0.4 | 1.6 |
The qwen3.8-max model charges one rate across its full 1M-token window. By contrast, qwen3-max steps from $1.2 to $2.4 to $3 per million input tokens as requests grow.
Global-scope deployments in Virginia, Frankfurt, Tokyo and Hong Kong list qwen3.8-max at $1.65 input and $4.951 output. Batch calls, where supported, cost 50% of real-time prices.
A limited-time 20% discount applies to qwen3.7-plus on top of its $0.4 and $1.6 list prices.
Is Qwen the Most Fine-Tuned Open Model on Hugging Face?
Hugging Face’s Summer 2026 report calls Qwen the community’s base model. Qwen-based models account for 151,448 derivatives on the Hub, against 82,506 for Google.
New Qwen derivative repositories arrived at roughly 180–210 a day through the first seven months of 2026. Outside developers built most of them: of 28,531 GGUF conversions of Qwen models, Qwen published 54.
Qwen GGUF files drew 39.6 million downloads a month, against 20.8 million for Gemma and 7.5 million for Llama. Llama-derived GGUF repositories slightly outnumber Qwen’s, per the report.
Across repositories with declared parameter counts, Qwen logged 2,045 million downloads in the 2026 window, against 37 million for Moonshot.
Small checkpoints carry most Hub volume, with models under 1B taking 83% of all-time downloads among those declaring a size, a split also tracked in these small language model adoption statistics.
Qwen vs Llama in the ATOM Report
| Model family | Cumulative downloads, Sept 2025 | Cumulative downloads, March 2026 |
|---|---|---|
| Qwen | 325.4M | 942.1M |
| Llama | 323.7M | 476.0M |
Qwen’s share of new fine-tunes and adaptations rose from 1% in January 2024 to 69% by February 2026, per the same report.
Meta’s share peaked at 44% in August 2024, then fell to 11%; more Llama data sits in these Meta AI usage statistics.
Alibaba’s December-quarter results gave a larger figure: more than 1 billion cumulative Qwen downloads on Hugging Face as of January 21, 2026. ATOM counts only mainline models, so the two totals cover different scopes.
167 Million Monthly Users: Qwen App Adoption in China
Alibaba says consumer-facing Qwen passed 300 million monthly active users across all platforms in February 2026. QuestMobile measures a narrower scope: the Qwen mobile app in China.
| AI-native app | MAU, March 2026 | MAU, June 2026 |
|---|---|---|
| Doubao | 345M | 382M |
| Qwen | 166M | 167M |
| DeepSeek | 127M | 129M |
QuestMobile recorded 5792.9% year-over-year growth for Qwen in its mid-year report. DeepSeek ranked third at 129 million; that rival’s figures are in these DeepSeek statistics.
Qwen’s citations of Bilibili content reached 1.705 million users on phone questions, 1.315 million on finance and 1.205 million on travel, per the same QuestMobile report.
Alibaba also tracks shopping through the app’s agents. About 140 million users had a first AI-driven shopping experience in Qwen app by the end of February 2026, and 250 million had since launch as of the August 20, 2026 results.
Alibaba’s Qwen AI Statistics on Revenue and Losses
Revenue from AI Cloud and Compute Services reached RMB48,437 million (US$7,139 million) in the quarter ended June 30, 2026, up 45%. AI-related product revenue hit RMB12,376 million, a twelfth straight quarter of triple-digit year-over-year growth.
| Metric (RMB million) | Q ended June 30, 2025 | Q ended June 30, 2026 | Reported YoY |
|---|---|---|---|
| AI Cloud and Compute Services revenue | 33,418 | 48,437 | 45% |
| AI Cloud and Compute Services adjusted EBITA | 2,419 | 5,628 | 133% |
| AI Labs and Applications revenue | 2,882 | 3,338 | 16% |
| AI Labs and Applications adjusted EBITA | (3,224) | (13,861) | (330)% |
| Capital expenditures | 38,676 | 67,678 | 75% |
The AI Labs and Applications segment groups AI model labs, Qwen Consumer Business Group and QwenWork. Alibaba attributes its wider loss to increased AI investment and higher inference cost for the Qwen app.
Capex drivers cited by Alibaba include procurement cycles, CPU-compute capacity for AI agents and higher chip component prices, a supply side covered in these AI hardware market share figures.
Omdia’s “AI Cloud Market: China – 2025” ranked Alibaba Cloud first with a 38.1% share. Buyer-side budgets appear in these enterprise AI spending statistics.
Where the Qwen AI Numbers Fall Short
Every Qwen3.8-Max score above comes from the Qwen team’s table, and harnesses differ by model. Qwen tested Terminal Bench 2.1 with Claude Code; other models show their best published score across harnesses.
Licensing tightened at the top end. Hugging Face notes that Qwen 3.8 2.4T added some non-commercial restrictions and revenue-share requirements.
Hub downloads leave out API usage, private deployments and other channels. Qwen also publishes weights on ModelScope, which Hub counts do not include.
User counts differ by scope. Alibaba’s 300 million covers all consumer platforms, while QuestMobile’s 167 million covers the Qwen app in China.
FAQs
How many people use Qwen AI?
Alibaba says consumer-facing Qwen passed 300 million monthly active users across all platforms in February 2026. QuestMobile counted 167 million monthly active users for the Qwen app in China in June 2026.
Is Qwen AI open source?
Many Qwen models ship open weights under Apache 2.0, per Hugging Face. The Qwen3.8 2.4T weights use a qwen3.8-max license that adds some non-commercial restrictions and revenue-share requirements.
How much does the Qwen API cost?
On Alibaba Cloud’s Singapore endpoint, qwen3.8-max lists at $2 per million input tokens and $6 per million output tokens as of September 24, 2026. Global-scope deployments list at $1.65 and $4.951.
Is Qwen more popular than DeepSeek?
In China’s app market, yes by QuestMobile’s count. Qwen had 167 million monthly active users in June 2026, against 129 million for DeepSeek. ByteDance’s Doubao led with 382 million.
Is the Qwen3.8 model discussed on Reddit the same as the Qwen3.8-Max API?
No. Qwen’s model card says the open Qwen3.8-2.4T-A95B checkpoint is text-only and always thinks. The hosted Qwen3.8-Max adds vision input, a non-thinking mode, built-in tools and a 1M-token default context.
Sources
https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
https://www.alibabacloud.com/help/en/model-studio/model-pricing
https://huggingface.co/blog/state-of-open-models-summer-2026
https://www.sec.gov/Archives/edgar/data/1577552/000110465926099220/tm2623667d1_ex99-1.htm
https://commandlinux.com/statistics/open-source-ai-model-downloads
