Note: '8-bit precision' badge is HF reading uint8 packing — model is 4-bit MXFP4

Correct base MTP acceptance: measured 64% @ MTP-3 (was pahajoki's 81% @ MTP-4)

Fix base_model: quantized FROM the distill (not regular Qwen)

Add benchmark.sh (reproduces the perf table)

Use measured base benchmarks (base 6k=85 not ~100); state parity precisely

Clarify MTP acceptance depths (grafted MTP-3 vs native MTP-4)

Document thinking mode is optional (verified both ways)

Add files using upload-large-folder tool

initial commit