Vision-OPD Qwen3.5-4B RandomDrop5 Step65

Qwen3.5-4B trained with Vision-OPD using random 5% visual-token retention for the student and full visual tokens for the teacher. This is the 65-step checkpoint.

The checkpoint can run with full visual tokens using standard Transformers. For 5% visual-token inference, use the pruning-aware serving code in prune-opd.

Training data: Vision-OPD-6K.

Downloads last month
11
Safetensors
Model size
5B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for zhuqiang/Vision-OPD-Qwen3.5-4B-RandomDrop5-Step65

Finetuned
Qwen/Qwen3.5-4B
Finetuned
(1,522)
this model