card: drop links to the private companion repo (file names kept as plain text)

Drop 'lossless' claim — replaced with 'deterministic / empirically within quantization noise' (reviewer correction)

Fix TL;DR table — HF renderer hides empty-header tables

Native MXFP4 AIME landed: 18/30 (NVFP4 +3 problems higher; greedy sampling variance)

Native MXFP4 MBPP landed: 0.921 / 0.783 (NVFP4 within noise)

Native HumanEval landed: 0.963 / 0.915 (NVFP4 within quantization noise)

Native MXFP4 GSM8K full: 0.9682 — NVFP4 within CI overlap

Relabel AIME as Non-Think greedy (was incorrectly labeled thinking=high)

Unify model card to canada-quant family template (H1 normalize, family table, absolute links, honest limitations, changes log)

Lead with throughput speedup vs upstream MXFP4 (the actual value prop)

Throughput vs upstream MXFP4 measured on same vLLM build (NVFP4 +7-37%)

Throughput section: apples-to-apples vs upstream MXFP4 (measurement in progress)

model card: fix stale queued line in TL;DR

v12: rewrite README with v12 shipping numbers, scrub historic debug-arc references

test

Add files using upload-large-folder tool

Add files using upload-large-folder tool

v0.3: BF16 mtp.0 + 5-patch install (MTP investigation findings + forward-compat framing)

v0.2: full quality suite (GSM8K 96.89%, MMLU-Pro 81.64%, HumanEval 95.1%/89.6%, AIME-30 66.67%, IFEval 84.84%) + throughput scaling + docker portability finding + e_proj forensic + GSM8K scoring fix

v0.1.1: add filed vLLM issues #43454, #43455 to upstream-contributions table

v0.1: model card with measured backend×format matrix and matched GSM8K

Add files using upload-large-folder tool

Add files using upload-large-folder tool

initial commit