Artificial intelligence · 2025

Qwen3

Alibaba's open family with switchable thinking, from 0.6B to 235B parameters.

Released in April 2025 under Apache 2.0, Qwen3 included dense and MoE models trained on about 36 trillion tokens, with a hybrid thinking mode. Qwen became one of the most widely fine-tuned base model families on Hugging Face and a staple of local inference.

See it in the living treeWikipedia