A compact multimodal model that handles both text and image inputs, quantized to 4-bit weights for reduced memory footprint while preserving much of the original model's capability. The w4a16 quantization scheme means it runs leaner than the full-precision version, making it practical on hardware with tighter memory constraints. Expect some trade-offs in precision compared to the unquantized original, but the core instruction-following and vision understanding remain functional.