A quantized variant of Qwen's 35B mixture-of-experts model, trimmed to roughly 4.75-bit precision to reduce memory footprint while preserving multimodal text and image understanding. The sparse activation pattern means only about 3B parameters fire per token, keeping inference costs manageable despite the large total parameter count. Trade-offs typical of aggressive quantization apply — slight degradation in precision-sensitive tasks compared to the full-precision original.