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

Upload README.md with huggingface_hub

Card: bundled mmproj-F32.gguf for vision

Card: add vision flags to the start command (--mmproj + --image-min-tokens 1024)

Card: best-of-both recommended, KL fidelity findings, BCB result, vision flags

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

Correct mmproj/cache claim: source says only --cache-reuse is disabled w/ multimodal (not 'cache reuse entirely'); rest unmeasured. (27B: + upstream-benchmark pointer)

Add upstream SWE-bench (Jackrong's base Qwopus-Coder, 67.0% thinking-off Q5_K_M) — clearly labeled NOT re-measured on this quant

f16 KV: state config only, no acceptance claim either way (pending more rigorous testing)

Correct unsupported claim: f16 KV is higher fidelity, NOT a measured acceptance lever (short-context A/B showed no difference vs q4_0)

Card: square corners on all tables (overflow:hidden + border-radius:0) + red critical-fork callout + amber section numbers

Card: datasheet/neo-brutalist redesign — monospace, hard square borders, numbered sections, symmetric spec grid, theme-adaptive (currentColor)

Card: styled HTML redesign (gradient header, pill badges, styled tables, GitHub alerts) — no collapsibles, theme-adaptive body, Jackrong-inspired

Card: vertical copy-paste launch command (one flag/line, \-continued, syntax-verified) + collapsible 'what every flag does' reference

Card redesign: centered header + badge bar, at-a-glance/files tables, collapsible launch/build/OpenCode, vision + perf sections (prototype for fleet-wide structure)

Card: add Vision section (--mmproj + --image-min-tokens 1024) and bundled chat_template.jinja note

Card: trim root-cause detail; document thinking-off cache fix (preserve_thinking) + OpenCode fork setup

Card: document agentic tool-calling fix (native tool_call:true), thinking on/off cache trade-off, update launch flags (n-max 5, cache-ram 64G)

provenance: set base_model_relation=quantized

Card: re-quantized from BF16 (genuine f16 emb); no imatrix (measured mixed on this coder)

Card: honest note on the f16 embeddings (Q8 source -> f16-typed but Q8 precision)

Add model card (two variants, off-thinking note, credits)