qwen38-mtp-head-4bit-cons

Qwen3.8-27B MTP head — 4-bit/g64, per-group best-of-two coding (no range shrink)

Requantisation of the organizer-pinned EigenLabs/Qwen3.8-27B-MTP-bf16 (@26a328e0) for the mlxfast Qwen-MTP challenge. Identical geometry and size to the naive mx.quantize(w, 64, 4) heads on the board (31 tensors, 238,934,093 bytes). Every group keeps its exact min/max endpoints — outlier anchoring is pointwise identical to the board's head. The only per-group choice is WHICH of two code-derivation rules is stored: the classic one (codes from fp32 scale, scale then rounded to bf16) or the self-consistent one (scale/bias rounded to bf16 first, codes derived from the stored values). The candidate with the smaller true as-stored reconstruction error wins, so by construction every group is at least as faithful as mx.quantize. 56% of groups switch; mean reconstruction error falls 5.1–7.4% per tensor with zero range shrink.