Qwen3.8-27B-BF16-SSMFIX-UD-Q3_K_XL

Source lineage:
  Qwen/Qwen3.8-27B
    -> redashes/Qwen3.8-27B-BF16-SSMFIX
    -> BF16 GGUF
    -> this GGUF

Quantizer:
  llama.cpp build 9222 (9a532ae4b)

Importance matrix:
  imatrix_unsloth.gguf_file

Equivalent llama-quantize arguments:

  --imatrix imatrix_unsloth.gguf_file
  --token-embedding-type q3_k
  --output-tensor-type q5_k
  --tensor-type attn_v=q5_k
  --tensor-type attn_gate=iq4_xs
  --tensor-type attn_qkv=iq4_xs
  --tensor-type ffn_down=iq4_xs
  --tensor-type ssm_alpha=iq4_xs
  --tensor-type ssm_beta=iq4_xs
  --tensor-type ssm_out=iq4_xs
  --tensor-type attn_k=iq4_xs
  --tensor-type attn_output=iq4_xs
  --tensor-type attn_q=iq4_xs
  --tensor-type nextn.eh_proj=iq4_xs
  <bf16-input.gguf> <output.gguf> iq3_s 12

Result:
  Quantization family: custom mixed UD-Q3_K_XL-compatible recipe
  Quantized size: 12807.91 MiB
  Reported rate: 3.93 BPW
  Output file size: 13441058848 bytes

The quantization log reports ssm_conv1d.weight tensors as F32. The eight SSMFIX source tensors are therefore retained in F32 by this recipe.