---
license: apache-2.0
---

<p align="center">
  <img src="https://huggingface.co/Winnougan/Klein-9b-kv-mxfp8/resolve/main/Flux2-Klein_00020_.png" width="720"/>
</p>

# Model Card for Klein-9B KV (MXFP8 / NVFP4)

Quantized KV-cache optimized variant of FLUX.2 Klein 9B for faster and more memory-efficient image generation.

---

## Model Details

### Model Description

Klein-9B KV (MXFP8 / NVFP4) is a quantized version of FLUX.2 Klein 9B with KV-cache support for improved performance in multi-reference and iterative workflows.

- **Developed by:** Black Forest Labs (base), Winnougan (quantization)
- **Model type:** Text-to-image / image-to-image generative model
- **License:** Apache-2.0 (repo), base model license applies
- **Finetuned from model:** FLUX.2-klein-9b-kv-fp8

---

### Model Sources

- **Repository:** https://huggingface.co/Winnougan/Klein-9b-kv-mxfp8

---

## Uses

### Direct Use

- Text-to-image  
- Image-to-image  
- Multi-reference workflows (KV-cache)

---

### Out-of-Scope Use

- Factual or reliable information generation  
- Safety-critical applications  

---

## Bias, Risks, and Limitations

- May reflect biases from training data  
- Minor quality loss due to quantization  
- KV benefits require compatible workflows  

---

### Recommendations

Use MXFP8 for better quality and NVFP4 for maximum performance. KV-cache is most effective in iterative workflows.

---

## How to Get Started with the Model

Load in a KV-compatible pipeline (e.g., ComfyUI) and reuse reference images to benefit from KV-cache.

---

## Training Details

### Training Data

Inherited from FLUX.2 Klein 9B. No additional training.

---

### Training Procedure

Post-training quantization:
- MXFP8  
- NVFP4  
- KV-cache enabled  

---

## Evaluation

### Results

- Faster inference with KV-cache  
- Reduced VRAM usage  
- Small quality trade-offs depending on format  

---

## Technical Specifications

- ~9B parameters  
- Rectified Flow Transformer  
- 1024×1024 resolution  

---

## Model Card Authors

Winnougan