End-to-end official BF16 → two EXL3 quants → Hugging Face → :12434 bench
with the existing 4.05bpw serve params and 8-bit KV.
- Plan (locked):
PLAN.md - Live log:
JOURNEY.md - Scripts:
scripts/
Source model: Qwen/Qwen3.8-Flash-Next.
ExLlamaV3: WORKSTATION/exllamav3.