Alibaba's Qwen3-30B-Thinking-2507 Hits 85 on AIME25 with Only 3.3B Active Parameters

On July 31, 2026, Alibaba's Qwen team released Qwen3-30B-A3B-Thinking-2507, a new reasoning model that quickly became one of the most talked-about open-weight releases of the year. The model has 30.5 billion total parameters, but only 3.3 billion are activated for any single answer. Despite this small active size, it scored 85.0 on the AIME25 math benchmark, beating many models that are ten times larger in active parameters.

The model uses a Mixture of Experts, or MoE, design. Think of it as a hospital with 128 specialist doctors, but only 8 walk into the room for any given patient. A small router network reads each token of input and decides which 8 experts are most useful for that token. The other 120 experts stay idle, so the model spends only a fraction of the energy a traditional dense model would spend.

This launch marks a clear turning point in the industry. For years, the AI race was about who had the largest model. In 2026, the race has shifted toward who can get the best results per active parameter. A 30B MoE model now matches or beats dense models above 70B on reasoning tasks, while running on a single consumer GPU instead of a full data-center rack.

For businesses, the practical impact is huge. Running this model costs only a few hundred dollars of electricity per month on a single mid-range GPU, compared to many thousands for a frontier model. That makes advanced reasoning available to small teams, local startups, and even individual developers for the first time.