Downloads · 30 days
7
23% of all-time downloads
francip/qwen38-mtp-head-q2-q4-mlp3-v1
qwen38-mtp-head-q2-q4-mlp3-v1 is a machine learning model from francip. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Qwen 3.8 27B MTP proposal head for the qwen3.8-27b-mtp-v1 track.
Downloads · 30 days
7
23% of all-time downloads
All-time downloads
31
Public
Parameters
121M
394 MB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors394 MB · 100%
How the weights are stored.
U3276.2M · 63%
From the Hugging Face model README
Qwen 3.8 27B MTP proposal head for the qwen3.8-27b-mtp-v1 track.
Derived from amal-david/qwen38-mtp-head-q2-q4-rerank-v1@ae6282749a52e052496dd5300b4aa441df7301e8
(affine-4/group-64 projections, affine-2 compact draft_lm_head, bf16 precision islands),
with ONE change: layers.0.mlp.{gate_proj,up_proj,down_proj} re-quantized at
affine-3 / group-64 directly from the organizer's bf16 head
(EigenLabs/Qwen3.8-27B-MTP-bf16@26a328e070875b0314d652a039b6b59902690f03) via
mlx.core.quantize(w, group_size=64, bits=3) (mlx 0.32.1). Every other tensor is
byte-identical to the parent artifact.
Purpose: 34 MB fewer weight bytes per draft step (134 -> 100 MB for the MLP) at acceptance measured neutral on mixed-difficulty fixtures. Proposal side only; the organizer-pinned target decides every emitted token.