Qwen3-30B-A3B

Qwen3-30B-A3B is Alibaba's Qwen3 generation mixture-of-experts model, released April 28, 2025 under the Apache 2.0 license. The model has 30.5 billion total parameters with approximately 3.3 billion active parameters per token, organized across 48 transformer layers with 128 experts and 8 experts activated per token using grouped-query attention. The native context window is 32,768 tokens, extendable to 131,072 tokens using YaRN position scaling. The model was trained on 36 trillion tokens—double the dataset of Qwen2.5—and supports 119 languages and dialects.

A key feature distinguishing Qwen3 from prior Qwen generations is a hybrid thinking/non-thinking mode: the model can switch between extended chain-of-thought reasoning (thinking mode, for mathematics, coding, and logic) and fast direct response (non-thinking mode, for general tasks) within a single deployment, controlled by an inference-time parameter. This dual-mode capability allows operators to tune the cost-quality trade-off without deploying separate models. The 30B-A3B MoE configuration competes with dense 70B-scale models due to expert routing efficiency while incurring approximately 3.3B-equivalent compute cost per token.

Qwen3-30B-A3B is available via Cloudflare Workers AI, AWS Bedrock, Fireworks AI, OpenRouter, Novita AI, and the Vercel AI Gateway. It is the smaller of the two Qwen3 MoE variants (alongside Qwen3-235B-A22B) and is appropriate for organizations needing strong reasoning and coding performance at lower per-token compute cost than the Qwen3-32B dense model.