Qwen3.6 arrives with new mixture-of-experts checkpoints
Originally announced by Qwen . We link the primary source so you can read it for yourself.
Alibaba’s Qwen team has released Qwen3.6, following Qwen3.5 by only two months. Reports indicate the update includes new open-weight checkpoints, among them a mixture-of-experts model in the family’s established efficient style.
What was released
The reported line-up continues Qwen’s pattern of offering both dense and mixture-of-experts options, the latter activating only a fraction of their parameters per token for faster generation, as in the existing Qwen3 30B-A3B. Specific sizes and the accompanying benchmark figures are the developer’s, and we treat them as claims pending independent results.
Why it matters
Two months between minor versions is a fast pace, and it is the pace itself that is worth noting. For people building on local models, a family that ships this frequently, under a permissive licence, is a moving target: it keeps improving, but it also means today’s best choice may be superseded quickly. That is a good problem to have, and a reason to design setups that make swapping models easy.
The context
This continues the cadence that has kept Qwen near the front of the open-weight field through 2026, alongside releases from DeepSeek, Mistral, and others. As with Qwen3.5, we will add specific Qwen3.6 entries to the model catalogue once we have confirmed their specifications and memory requirements against primary sources, rather than launch-day figures.