Alibaba's Qwen family of AI models has surpassed 3 billion downloads, overtaking the open-model offerings of Meta and Google to become the most downloaded open AI models in the world, Bloomberg reported on August 15, 2026.
The milestone caps a remarkable run for the Chinese e-commerce giant's open-weight strategy, which has seen Qwen releases accumulate adoption at a pace that now dwarfs Western rivals. For developers and enterprises following the open-source AI race, the shift marks one of the clearest measurable reversals in the global AI landscape.
What the numbers show
Bloomberg's report, picked up by Business Standard and BusinessLine, puts Qwen's cumulative downloads above 3 billion across model hubs, ahead of Meta's Llama family and Google's Gemma models. Platform data reviewed by AI Buzz Wire underscores the current velocity: on Hugging Face alone, the top 100 Qwen repositories recorded roughly 330 million downloads in the past 30 days, with the compact Qwen3-0.6B leading at more than 30 million monthly downloads.
The download crown matters because open-weight models compete primarily on adoption. Developers who build on a model family create an ecosystem of fine-tunes, tools and derivatives that in turn attracts more users — a flywheel Meta pioneered with Llama and that Qwen has now decisively claimed.
Downloads are an imperfect proxy for usage — a single developer may pull many models, and production deployments are invisible in the count — but they remain the industry's standard yardstick for open-weight traction. Hugging Face downloads in particular have become a benchmark that Meta itself cited when Llama ruled the category, which makes Qwen's overtaking of it symbolically hard to dismiss.
The Qwen 3.8 release driving the surge
The latest milestone lands just days after Alibaba's Qwen team released the open weights for Qwen 3.8 under the Apache 2.0 license, according to The Decoder.
The core release, Qwen3.8-27B, is a multimodal dense model with 27 billion parameters that Qwen says outperforms the larger Qwen3.7-Plus on coding and office tasks. The team highlights improved agent capabilities, with the model planning more independently and completing tasks more reliably. It natively handles up to 262,000 tokens of context and can scale to one million using the YaRN method, while processing images and video — including diagrams, documents and multi-hour footage. A flexible thinking mode is enabled by default but can be toggled per query.
Alibaba also shipped weights for the much larger Qwen3.8-2.4T-A95B, a sparse model built to operate at the Max tier that the company has positioned against frontier competitors. Both models are available on Hugging Face and ModelScope, and a hosted version with one million tokens of context is coming to Qwen Cloud, Alibaba's AI service.
The Max-tier model extends the family's reach into frontier territory. According to The Decoder's coverage, the open-weight Qwen3.8-Max is a 2.4-trillion-parameter system designed for long-horizon AI tasks, and Alibaba has positioned it as a rival to leading Western models including Anthropic's Claude. Releasing weights at that scale under a permissive license is unprecedented: until recently, models in the trillion-parameter class were the exclusive property of a handful of American labs.
From free weights to paid tiers
The download record does not mean Alibaba is giving everything away. The company has signaled that large-scale users of future Qwen models will face charges, and its flagship releases increasingly pair open weights with paid hosted services — a playbook that treats openness as distribution rather than charity.
That strategy sits against a broader backdrop. A Stanford University analysis found that China captured the global lead in open-weight AI development during 2025, with Chinese models accounting for a growing share of top-performing releases that ship with downloadable weights. Qwen's 3 billion downloads are the most visible proof point of that shift.
Why Meta and Google lost the lead
Meta pioneered the modern open-weight movement with Llama in 2023, but its commitment has wavered as the company chases frontier-scale closed models. Google, meanwhile, releases Gemma models as distilled companions to Gemini rather than as its primary strategy. Neither has matched Qwen's cadence — Alibaba's team ships new model families and sizes at a pace that keeps the ecosystem refreshed.
Alibaba has also benefited from distribution most Western labs lack. Qwen models ship not only through Hugging Face but through ModelScope, Alibaba's own model hub, which dominates in China and across Southeast Asian developer communities. Every release is accompanied by ready-to-run fine-tunes, quantized variants and documentation in multiple languages, lowering the barrier for adoption in markets where Meta and Google have thinner roots.
For developers, the calculation is pragmatic: the most capable openly downloadable models now come from a Chinese company, and licensing under Apache 2.0 makes commercial use straightforward. For policymakers in Washington and Brussels, Qwen's milestone is likely to intensify debates over whether restricting Chinese open-weight models is even feasible when they already anchor so much of the world's downstream AI development.
Stay Ahead of AI Read more AI news →
