AntSpeaker is the official repository for MECT (Mixture of Experts with CNN-Transformer Network), a speaker verification model from Ant Research described in arXiv paper 2609.24061. MECT integrates a Mixture-of-Experts mechanism into a CNN-Transformer backbone with an optimized block structure and stacking scheme, and the paper explores four MoE variants that span utterance-level and frame-level granularity with dense and sparse routing. The checkpoints are hosted on Hugging Face and come in four sizes, from 3.78M parameters (MECT-A1) to 9.57M (MECT-B2), for six files in total: A1, A2, B1 and B2 trained on VoxCeleb2, a B2 trained on VoxCeleb2 plus VoxBlink2, and a causal B2 variant intended for streaming speaker verification. On the VoxCeleb1 test sets, the reported EER for VoxCeleb2-only models ranges from 0.27% (B2, Vox1-O) to 1.05% (A1, Vox1-H). The B2 model trained on VoxCeleb2 plus VoxBlink2 reports 0.23% / 0.29% / 0.54% EER on Vox1-O / Vox1-E / Vox1-H, and 0.22% / 0.28% / 0.52% with calibration applied. Usage is a short PyTorch example: load a checkpoint through the antspeaker package, compute 80-dimensional Mel filterbank features with torchaudio, extract an embedding, and compare two utterances by cosine similarity. The repository contains the antspeaker model code, an example script, and a requirements file; the checkpoints are downloaded separately from Hugging Face. The code and weights are released under CC-BY-NC-SA 4.0, which restricts use to non-commercial purposes and requires share-alike, so commercial deployment would need separate permission from the authors.