Geometric-AI/DeepSeek-V4-Flash-0731-ROCmFP3-MIX

card: whole-file bits-per-weight (2.766 vs reference 2.88 at equal score)

card: publish both serving configs with measured quality/speed tradeoff

card: clean v3-only read

card: v3 leads (98.29 GB, 17/17 + 82/92 held-out, 96GiB window)

add v1's REQUIRED gumix sidecar (was missing)

v3: 98.29 GB, 17/17 COMPSEC + 82/92 held-out, fits Strix 96GiB window

Upload README.md with huggingface_hub

docs: warn against --ds4-expert-top-k 4 on this adaptive artifact

fix: correct the published sha256 (was a pre-release build's hash)

Rebuild rotation-free: the decoder does not implement rotation, so the previous file could not load

Option 1: qtype-105 adaptive down-experts, codebooks embedded in the GGUF

initial commit