A mid-sized mixture-of-experts model from the Qwen3.6 family, quantized to INT8 precision using W8A8 quantization — meaning weights and activations are both stored at 8-bit, reducing memory footprint while aiming to preserve output quality. It accepts both text and image inputs, making it multimodal. The quantization trade-off is typical: slightly reduced precision in exchange for faster inference and lower hardware requirements.