Downloads · 30 days
0
qwen38-mtp-head-fcbf16
Cyleux/qwen38-mtp-head-fcbf16
qwen38-mtp-head-fcbf16 is a machine learning model from Cyleux. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Mixed-precision repack of the pinned MTP head EigenLabs/Qwen3.8-27B-MTP-bf16 (@26a328e0) for the Layr-Labs qwen-3.8-mtp-challenge declared-head surface: fc.weight kept in original bf16 (per-tensor session measurements…
Access
Public
Updated Aug 17, 2026
Parameters
111M
314 MB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors314 MB · 100%
Parameter types
How the weights are stored.
BF1664.1M · 58%
Base models
Model card
From the Hugging Face model README
Qwen 3.8 27B MTP head — fc in bf16, rest 4-bit group-64
Mixed-precision repack of the pinned MTP head EigenLabs/Qwen3.8-27B-MTP-bf16
(@26a328e0) for the Layr-Labs qwen-3.8-mtp-challenge declared-head surface:
fc.weight kept in original bf16 (per-tensor session measurements attribute
the entire draft-acceptance cost of uniform 4-bit quantization to this one
tensor), the seven other 2D projections at affine 4-bit group-64 (MLX 0.32.0
mx.quantize, the geometry of the current frontier head), 1D norms bf16.
Every kernel this head dispatches is already exercised by promoted runs:
bf16 GEMV by the pinned head, 4-bit group-64 by the frontier head and the
backbone. The head only proposes — the pinned target verifies every emitted
token.
- License
- apache-2.0
- Created
- Aug 17, 2026
- Updated
- Aug 17, 2026