KAT-Coder-V2.5-Dev — bmk asymmetric quantization (IQ2_XXS/Q2_K experts, Q8_0 dense)
Asymmetric GGUF quantization of Kwaipilot/KAT-Coder-V2.5-Dev (qwen35moe: 34.7B total / ~3B active, 40 layers, 256 routed experts top-8, 1 shared expert, GDN hybrid attention, 262144 ctx), built for the bmk bare-metal inference engine.
Quantization map
| Component | Type |
|---|---|
| Routed experts gate/up (40 × 256) | IQ2_XXS |
| Routed experts down (40 × 256) | Q2_K |
| Attention incl. GDN/SSM, shared experts, embeddings, output head | Q8_0 |
Importance-matrix calibrated (Kwaipilot imatrix, 510 entries / 802 chunks). File size: 10.9 GiB. Designed for 8/12/16 GB VRAM cards via expert-granular streaming (only the top-8 routed experts per layer are loaded).
Usage
Run with bmk:
bash download_model.sh bmk-q2
LD_LIBRARY_PATH=vendor/cuda-runtime ./bmk -m models/kat-coder-asym.gguf -p "..." -n 64
llama.cpp-compatible GGUF; the tensor layout follows the standard qwen35moe architecture.