This is Qwen3.8-27B-Unleashed-GGUF
made by outsourc-e with fixed by me scale for ssm_conv1d.weight tensors:
| Tensor | QType | α/δ/σ̄ | S_b | S_a |
|---|---|---|---|---|
| blk.52.ssm_conv1d.weight | F32 | α=0.59005 | 0.0016 | 0.0006 |
| blk.53.ssm_conv1d.weight | F32 | α=0.55484 | 0.0015 | 0.0005 |
| blk.56.ssm_conv1d.weight | F32 | α=0.54486 | 0.0015 | 0.0004 |
| blk.57.ssm_conv1d.weight | F32 | α=0.53574 | 0.0015 | 0.0004 |
| blk.58.ssm_conv1d.weight | F32 | α=0.60972 | 0.0012 | 0.0005 |
| blk.60.ssm_conv1d.weight | F32 | α=0.48136 | 0.0017 | 0.0004 |
| blk.61.ssm_conv1d.weight | F32 | α=0.65327 | 0.0010 | 0.0004 |
| blk.62.ssm_conv1d.weight | F32 | α=0.61856 | 0.0013 | 0.0005 |
Genesis SVD with Noise Gate supression can't be applied to this model, because it can process only BF16, F32, MXFP4, Q8_0 and Q4_0 tensors. NVFP4 encoding is not supported in gguf-py in llama.cpp. I can read NVFP4 weights, but can't modify them. So I fixed only 6 too loud ssm_conv1d.weight layers in this model. Everything else left unchanged.