The Open-Source AI Stack
RSS

The Stack · Weights · Open source

Qwen (Alibaba)

Alibaba's aggressive open-weights series (Qwen 2.5 / 3); Apache 2.0 across most sizes; full-precision weights available.

Apache 2.0 · stable · Project site →

Qwen is Alibaba Cloud's model family. Qwen 2.5 (late 2024) and Qwen 3 (2025) shipped a wide range of sizes (0.5B to 72B+) under Apache 2.0, including specialty variants for code, math, and very-long-context. Qwen has been notably aggressive at the open-permissive end: full-precision weights, Apache 2.0 across most sizes, no field-of-use restrictions like Llama's MAU clause. Qwen matters because it has emerged as the most-downloaded open- weights family in 2025-2026, displacing Llama at the top of HuggingFace's monthly download counts in many months. The capability is competitive with Llama and Mistral peers at the same size class. Compared to siblings: Llama (similar size coverage but the Community License is restrictive), DeepSeek (MIT, frontier-class reasoning, similar open posture but smaller release cadence), Mistral (early Apache 2.0 then drift), OLMo (truly open including data, smaller capability ceiling). Qwen's distinctive angle is "Apache 2.0 across the family including the flagships, with strong math and code variants." Production-ready and widely deployed in 2026. Used as the base for many fine-tunes on HuggingFace; the long-context Qwen variants are popular for retrieval-heavy use. Strategic question: does Qwen's open posture survive future regulatory pressure on Chinese AI exports, and does the West's "do we trust models from a Chinese lab" debate affect adoption.

Sources

Want a follow-up? Ask the chat about Qwen (Alibaba) in context. It will compare to siblings at the same layer and ground every claim in the wiki.

Other projects at the Weights layer

9 siblings · ordered open first