---
title: Nucleus-Image-NVFP4_mixed
canonical_url: "https://www.modelscope.cn/models/ApacheOne/Nucleus-Image-NVFP4_mixed"
md_url: "https://www.modelscope.cn/models/ApacheOne/Nucleus-Image-NVFP4_mixed.md"
repository: ApacheOne/Nucleus-Image-NVFP4_mixed
last_updated: 2026-04-19
license: apache-2.0
model_type:
  - custom
base_model:
  - NucleusAI/Nucleus-Image
base_model_relation: quantized
parameters: 25.7B
tensor_type:
  - U8
  - F8_E4M3
  - F32
  - BF16
library_name:
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 19
stars: 0
---

# Nucleus-Image-NVFP4_mixed

> Nucleus-Image-NVFP4_mixed - ApacheOne 在 ModelScope 开源的模型。I am still getting OOM on a L4 24GB , 64gb ram system.

ApacheOne/Nucleus-Image-NVFP4_mixed 是 ModelScope 魔搭社区上的 25.7B 参数机器学习模型，采用 apache-2.0 许可，基于 NucleusAI/Nucleus-Image 构建。

- **Repository**: ApacheOne/Nucleus-Image-NVFP4_mixed
- **License**: apache-2.0
- **Parameters**: 25.7B
- **Base model**: NucleusAI/Nucleus-Image
- **Downloads**: 19
- **Stars**: 0
- **Last updated**: 2026-04-19

Source: https://www.modelscope.cn/models/ApacheOne/Nucleus-Image-NVFP4_mixed

---

# UPDATE 

I am still getting OOM on a L4 24GB , 64gb ram system.

Aggressive quant is most likely a failure and will not work as is. 

2d weights only quant does load and run but still getting OOM from improper loading method. 


Not sure of the best way to even run the base model yet, Should work where you offload the dense layers into ram and only leave the active layers in vram. 

# UNTESTED

I cant test this model myself as its too big. 

`Nucleus-Image_noreshape2Dweightonly.safetensors` This model has only the 2D non-MoE expert layers quantized to nvfp4 which should give speed to the active layers in vram without loss of the dense layers. 

I will try testing this model once more support comes out for the loading methods of the model.  

`Nucleus-Image_transformer_aggressive_nvfp4.safetensors` This model is every layer quanted just for context reasons of what is possblie. The dense layers might be sensitive to the quant.