Views
No views yet
| file | size | |
|---|---|---|
KAT-Coder-V2.5-Dev-MTP-APEX-i-quality.gguf | 20.72 GB | recommended — includes a working MTP head for speculative decoding |
KAT-Coder-V2.5-Dev-APEX-dynamic.gguf | 11.86 GiB | sized for a 16 GB card |
kat-coder.imatrix | 192 MB | the importance matrix, reusable for your own tiers |
model-00014-of-mtp.safetensors | 1.69 GB | the bf16 MTP head, if you want to redo the transplant |
mtp_num_hidden_layers: 0 — no MTP head at all, so no speculative decoding is
possible out of the box, and that is true of the vendor release and of every other quant of this
model I am aware of. The build above transplants Qwen3.6-35B-A3B's own trained head onto it:
2.03× on a hard agentic-coding suite with correctness unchanged (100% on both suites either
way — the head only drafts, the main model verifies).