Repoint build instructions to charlie12345/ROCmFPX (ROCmFPX FP3/4/6/8 repo)

Upload README.md with huggingface_hub

consolidate to single best-balance rocmfp4 quant

consolidate to single best-balance rocmfp4 quant

consolidate to single best-balance rocmfp4 quant

consolidate to single best-balance rocmfp4 quant

Upload README.md with huggingface_hub

Upload README.md with huggingface_hub

Add mmproj-F32.gguf (best-of-both / vision)

Add Qwen3.6-40B-Deckard-MTP-ROCmFP4-COHERENT-embQ8-headQ6.gguf (best-of-both / vision)

Add Qwen ASCII logo header (centered logo+title, full-width spec table)

Deckard: neutralize last 'off-thinking coding' flag wording (config description only)

Claim audit: remove cross-model generalizations (Q6-head %, f16-draft-head, decode comparisons) — state per-model config only, no borrowed sibling measurements

Deckard: soften unverified 'does better off-thinking' claim — present off-thinking as an option, attribute the trend to the Qwen3.6 coder family (cited), note it's not measured on Deckard

Card: datasheet/neo-brutalist restyle (monospace, hard borders, spec grid, numbered sections); claims audited (f16 KV stated as config only)

Card: note bundled chat_template.jinja (froggeric unified template) for tool calls + inline think-toggle + vision

Add froggeric v20 fixed chat template (vision + inline think-toggle + tool anti-stall); use via --chat-template-file

Card: add Vision section (Qwen3-VL mmproj + --image-min-tokens 1024 fix for incorrect image reads)

provenance: set base_model_relation=quantized + repoint base_model at actual source

Card fix: imatrix variant sizes are 21.8GB (4-bit head) / 22.2GB (Q6 head), not both 21GB

Drop mtpF16 (f16 MTP draft head measured as a wash vs 4-bit)

Card: add imatrix variants + measured KL/PPL; mark imatrix recommended; drop mtpF16

Add imatrix headQ6 variant (NEO-style imatrix + Q6_K output head)

Add imatrix base variant (NEO-style general+code imatrix; measured -10% median KL, +0.8pp top-token vs non-imatrix)

Card: controlled test — f16 MTP head is a wash (~0.83 both); mtpF16 redundant, use base

Card: f16-head warm acceptance ~0.85 sustained (cold 0.77 undersold it); not an underperformer

Card: honest f16-head result (did NOT beat 4-bit Qwopus-donor head; donor match > precision)

Card: document the f16-MTP-head variant (3 variants)

Add f16-MTP-head variant (BF16 27B donor head kept at f16; draft acceptance rate = 0.76852)

Add model card (lineage, functionally-verified status, block_count gotcha)

Add Q6-head variant

Add base variant (ROCmFP4 STRIX + Q8 emb; transplanted MTP head)

initial commit