☀️ AI Morning Minute: Qwen
Alibaba built one of the most downloaded AI model families in the world. Most people in the West still know nothing about it.
The conversation about frontier AI tends to center on OpenAI, Anthropic, and Google. That framing is missing something significant. Qwen, the family of AI models built by Alibaba’s Tongyi Lab, has quietly become the most downloaded open-weights model family on the planet, with over 200,000 derivative models built on top of it and a consumer app that reached 234 million users by May 2026.
What it means
Qwen (pronounced “chwen,” from the Chinese 通义千问, meaning “a thousand questions”) is Alibaba’s series of large language models, ranging from tiny edge models that run on a phone to flagship models with hundreds of billions of parameters. The family handles text, images, audio, and video, supports over 119 languages, and is released under the Apache 2.0 license, meaning anyone can download, modify, and build commercial products on top of it without paying Alibaba.
The current generation, Qwen 3.6 and 3.7, includes a standout feature: hybrid thinking mode. The same model can switch dynamically between fast, cheap outputs for simple questions and slow, step-by-step reasoning for complex ones. You don’t pick a “smart” model or a “fast” model. One model reads the task and decides which mode fits.
By April 2026, Qwen model downloads had approached 1 billion, accounting for over 50% of all open-source model downloads globally. Airbnb’s customer service chatbot runs largely on Qwen models. That’s not a pilot.
Why it matters
The benchmark numbers are no longer easy to dismiss. Qwen3.5-Max-Preview sits in the top five on LM Arena globally, first among Chinese models. Qwen 3.6-Plus matches Anthropic’s Claude Opus 4.5 on SWE-bench and Terminal-Bench 2.0 programming benchmarks. A year ago those comparisons would have been charitable. Now they’re accurate.
Export controls haven’t stopped the model from spreading. The U.S.-China Economic and Security Review Commission specifically noted that open-weights models like Qwen have been critical to China’s ability to develop AI capabilities despite chip restrictions. Open weights travel. Once a model is released publicly, access controls on hardware become much harder to enforce.
It’s cheaper to run than Western alternatives. Qwen’s inference costs through Alibaba Cloud run significantly below comparable closed models from OpenAI and Anthropic, and the open-weights versions can be self-hosted for essentially the cost of compute. For price-sensitive markets and enterprise deployments outside the U.S., that matters more than brand recognition.
Simple example
Let’s say that you need a capable AI model for your company’s internal tools. You don’t want your data leaving your servers, you need multilingual support, and you can’t afford $50 per million tokens at scale. You download Qwen 3.6, run it on your own infrastructure, and pay nothing to Alibaba. The model handles your documents in English, Spanish, and Mandarin. Nobody outside your building sees the data.
But there is a catch: Qwen was built by a Chinese company, trained on Chinese infrastructure, and subject to Chinese law. What’s in the training data, what values shaped its outputs, and what backdoors if any exist are questions you can inspect the weights for but not fully answer. Open weights means you can see the model. It doesn’t mean you can see everything that went into making it.