Qwen3-30B-A3B is the small Mixture-of-Experts model released with the Qwen 3 generation on 29 April 2025. It holds 30.5 billion parameters in 128 experts across 48 layers, but routes each token through only 8 of them, so roughly 3.3 billion parameters are active at any moment. That ratio is the whole point of the model: it answers with the breadth of a large network while costing about as much to run as a small one, which is why it became the size most people actually self-host.
The model carries the Qwen 3 hybrid switch between a thinking mode for maths, code and logic and a direct mode for ordinary dialogue, inside one set of weights. Context is 32,768 tokens natively and 131,072 with YaRN scaling, which the model card describes as a setting to enable only when long inputs are genuinely needed. Over 100 languages and dialects are supported, function calling and agent frameworks included.
Alibaba published the weights under Apache 2.0, commercial use included, and the repository has drawn around 1.9 million downloads in the past thirty days — a scale of adoption matched by few open models of any size. Two refreshed editions under the same name followed later in 2025 (an instruction variant in July with a far longer context window, and a reasoning variant in August); this profile documents the original April release, which remains the base of countless third-party fine-tunes.