DeepSeek-R1-W4AFP8

This model is a mixed-precision quantized DeepSeek-R1, with dense layer using FP8_BLOCK_SCALING, MoE layers uses INT4 weights and FP8 activation.