Qwen and the rise of open-source AI from China

If you want to understand why open-weight AI became a genuine rival to the closed frontier, start with Qwen — the model family from Alibaba's Qwen team that is, by most counts, the most-downloaded open AI in the world. The story of how it got there is also the story of the open-weight era's central strategic bargain: give away enough to become default infrastructure, then decide what to charge for when the frontier gets expensive.

Qwen is a broad family of models — text, code, vision, image, and increasingly audio — most released under the permissive Apache 2.0 licence, meaning anyone can download, run, fine-tune and commercialise them without royalties or usage fees. That openness, shipped at a relentless cadence and across a wide range of model sizes (0.6B to 235B parameters in the Qwen3 generation), is what built the ecosystem. By 2026, Qwen is reported to account for over half of all open-source model downloads, approaching a billion in total — a scale that puts it alongside Linux or PostgreSQL in the 'became infrastructure by being free' category.

The architecture has matured across generations. Qwen3.6 — the April 2026 release that sharpened the family's positioning — uses Mixture-of-Experts (MoE) sparse activation in its larger models, meaning compute scales more efficiently than a dense model of equivalent parameter count. The flagship open model, Qwen3.6-27B (released April 2026), is dense, fits in roughly 16.8 GB at 4-bit quantisation so it runs on a single consumer GPU, and on Alibaba's own benchmarks outperforms the much larger 397B predecessor on agentic coding tasks (77.2% vs 76.2% on SWE-bench Verified; 59.3% vs 52.5% on Terminal-Bench 2.0). Across the range:

The open models cover most production use-cases. The closed one sits at the frontier.

The Hugging Face download numbers tell the structural story. Open models become infrastructure precisely because the barrier to trying them is zero — a developer can pull a Qwen checkpoint, fine-tune it on proprietary data, and ship it in a product without a licence conversation or a per-token bill. By April 2026, MySummit reports, Qwen accounted for over half of all open-source model downloads worldwide, approaching a billion in total, with 433 open models published for download. Alongside it, DeepSeek, Moonshot AI's Kimi models, and dozens of specialised code, reasoning and multilingual models build on Qwen foundations or compete directly with Qwen for that base-model position. The open-weight wave is not one company — it is an ecosystem, and Qwen is the substrate much of it runs on.

Pure openness has limits when you need revenue. On 20 April 2026, Alibaba released Qwen3.6-Max-Preview as its first closed-weight, API-only flagship — the most capable Qwen model is now behind a paywall, accessible only through Alibaba Cloud's Model Studio and DashScope endpoints ($1.30/M input, $7.80/M output tokens). No Hugging Face download. No ModelScope release. No self-hosting. For the first time in Qwen's three-year history, the best model is not in the community's hands. The smaller open models — Qwen3.6-27B and the 35B MoE — kept shipping under Apache 2.0, so the family is still far more open than anything from OpenAI or Anthropic. But the direction of travel is clear: openness is the on-ramp, and the frontier is where Alibaba collects.

This is not unique to Qwen. It is the central tension of the open-weight era: every major open-weight lab eventually faces the same calculation. Training frontier models costs hundreds of millions of dollars in compute. The open strategy wins ecosystems but does not directly fund those bills. At some point the most capable model has to go behind a meter. Qwen3.6-Max-Preview is that moment for Alibaba — earlier than some expected, later than critics feared, and perfectly predictable in retrospect. Read alongside DeepSeek, Kimi and MiniMax, Qwen is both the leader of the open wave and the first to openly test where its ceiling is.

The question that matters now is whether the open tier stays genuinely competitive, or gradually becomes the demo that pushes developers toward the paid frontier. So far the open models are real and capable — Qwen3.6-27B outperforming a far larger predecessor is a meaningful benchmark claim — but the pattern to watch is the gap between the open and closed tiers over successive generations. If that gap widens deliberately, the open tier shifts from a product to a funnel. If it stays narrow, Qwen's openness is durable. That is the metric worth tracking, not the download count.