Qwen3 ☆

Qwen3 (April 2025) spans dense models from 0.6B to a 235B/22B-active MoE, all Apache-2.0, with a controllable thinking mode and strong multilingual and coding results. The Qwen family passed 100,000 derivative models on Hugging Face — the most-built-upon open line in the world — making Alibaba the quiet default of global open-source AI. Qwen3-Max (closed, 1T+ parameters per company statements) sits above it.