A compact mixture-of-experts variant from the Qwen 3 family, this model activates only a subset of its parameters per token, keeping inference costs lower while retaining broad language understanding. It handles long contexts well, with a 262K token window that suits document-heavy tasks. As a tiny-random publish, it's a community repackaging rather than an official release, so behavior may differ slightly from canonical Qwen 3 MoE checkpoints.