A mid-sized mixture-of-experts model that activates only 3 billion of its 35 billion parameters per forward pass, keeping inference costs lean while retaining a broad parameter pool for diverse tasks. It handles both text and image inputs, making it multimodal out of the box. The FP8 dynamic quantization trades a small amount of numerical precision for significantly reduced memory footprint, which shows up as slightly lower fidelity on very fine-grained reasoning tasks.