Key Info
DeepSeek Flash's expert layers are already natively quantized to FP4, according to dax, so attempts to quantize the model further bring minimal gains.
Highlights
- The expert weights in DeepSeek Flash ship in FP4 precision natively.
- Additional quantization of the model offers little practical benefit.
- dax questions the common claim that people are "quantizing DeepSeek," calling it confusing given the model's existing precision.