Downloads · 30 days
0
qwen38-mtp-head-4bit-mse
Lie24/qwen38-mtp-head-4bit-mse
qwen38-mtp-head-4bit-mse is a machine learning model from Lie24. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Requantisation of the organizer-pinned EigenLabs/Qwen3.8-27B-MTP-bf16 (@26a328e070875b0314d652a039b6b59902690f03) for the mlxfast Qwen-MTP challenge.
Access
Public
Updated Aug 17, 2026
Parameters
66.4M
239 MB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors239 MB · 100%
Parameter types
How the weights are stored.
U3253.1M · 80%
Model card
From the Hugging Face model README
Qwen3.8-27B MTP head — 4-bit/group-64 affine, error-minimising quantisation
Requantisation of the organizer-pinned EigenLabs/Qwen3.8-27B-MTP-bf16
(@26a328e070875b0314d652a039b6b59902690f03) for the mlxfast Qwen-MTP challenge.
Identical geometry to the naive mx.quantize(w, group_size=64, bits=4) heads
already on the board (8 quantised 2D matrices + 7 bf16 norms, 31 tensors,
238,934,093 bytes, single bare model.safetensors) so the scales-keyed loader
walk needs no code change. The difference is only in how each group's
(scale, bias) is chosen: a per-group search over a symmetric shrink of the
min/max range, scored by the reconstruction error AFTER scale/bias are rounded
to bf16 (the precision they are stored and read at).
- mean relative-L2 dequantisation error reduced 5.72% vs
mx.quantize - file sha256
10670a03c470db9d8384ed543b3cae0e921d34fcc324c83a0c512e08e1755650 - runner tree digest
60e6dd6d19edfb48606c532b6977d5960f32ad588fe53e54d8d04a3481d19503
The head only proposes tokens; the organizer-pinned target re-verifies every emitted token against the hidden serial trajectory.
- Created
- Aug 17, 2026
- Updated
- Aug 17, 2026