DEEPSEEK 1.58BIT GGUF (HUGGING FACE HUB)

Summary from the TLDR AI newsletter

The unsloth folks have figured out which piece of the new R1 model can be properly quantized. They also found some tokenizer quirks to be aware of, which make quantization slightly harder. In summary, just the MoE layers go to 1.58 bit while everything else remains in 4 or 6 with their dynamic quantization scheme.