command-center/agent/memories/sharpshooter/wt-20261017-173221-6d1fc1f-cyy12v5yf03k/architecture.md

Model Format

• Quantization — established across all LLM endpoints; each service (radiance, halogen-flash-server, llama.cpp) applies its own quantization scheme as part of the runtime.
• Qwen3.8-Flash-Next .rad — radiance container uses .rad format natively; the format encodes routed experts as 4-bit codes into a 16-level non-uniform codebook (128-pt Walsh–Hadamard rotation, GPTQ on calibration + 3 coordinate-descent sweeps, E4M3 scales per 64 weights), with 10 whole bf16-protected experts/layer, int8 W8A8 trunk/lm_head, fp8 HC mixtures, and fp8 fixed-scale PLE n-gram.