Its 256k context window. Quantization is orthogonal. We cant really tell directly so it could be quantized.

The model is already natively MXFP4-quantized during training, so there is no quality loss.