Downloads · 30 days
25
40% of all-time downloads
Qwen3.5-4B-NQ-gdnw8
JD97/Qwen3.5-4B-NQ-gdnw8
Qwen3.5-4B-NQ-gdnw8 is a machine learning model from JD97. 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.
JD97/Qwen3.5-4B-NQ 기반. GDN linearattn 입력 projection(inprojqkv, inprojz)을 W4→W8로 bump해 GPQA razor-thin 미달(NQ full GPQA=0.6212, gate 0.630) 회복을 노린 변형.
All-time downloads
63
Public
Parameters
5.3B
5.4 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors5.4 GB · 100%
Parameter types
How the weights are stored.
I323.9B · 74%
Base models
Model card
From the Hugging Face model README
Qwen3.5-4B-NQ-gdnw8 (mixed-precision recovery variant)
JD97/Qwen3.5-4B-NQ 기반. GDN linear_attn 입력 projection(in_proj_qkv, in_proj_z)을 W4→W8로 bump해 GPQA razor-thin 미달(NQ full GPQA=0.6212, gate 0.630) 회복을 노린 변형.
구성 (NQ 대비 변경점만)
- group_0: W4 g32 (Linear body: MLP, GDN out_proj)
- group_1: W8 g32 (linear_attn.in_proj_qkv, in_proj_z) ← bumped (SSM 입력 projection, 민감)
- group_2: W8 g32 (lm_head, NQ와 동일)
- ignore(BF16): linear_attn.in_proj_a/b, full-attn q/k/v/o(8층), vision, MTP, embed
- RTN(data-free) 양자화. 멀티모달 구조 보존.
서빙
NQ와 동일하게 docker/lmhead-quant.patch(W8 lm_head용 패치 vLLM) 필요. docker/serve.py 참조.
근거
KL forward-only sensitivity 연구(SSM-Transformer 하이브리드)상 SSM 입력/cross-projection이 가장 민감, decay(dt/a-b)는 둔감. NQ가 a/b는 BF16 보존하면서 민감한 in_proj는 W4로 깎은 배분을 교정.
- Type
- qwen3_5
- License
- apache-2.0
- Created
- Jun 18, 2026
- Updated
- Jun 18, 2026