MN-Violet-Lotus-12B-NVFP4

MN-Violet-Lotus-12B NVFP4 Quant

This is an NVFP4 quantized version of FallenMerick/MN-Violet-Lotus-12B.

The original model is a high-quality merge optimized for roleplaying, creative writing, story generation, and strong emotional intelligence (EQ benchmark ~80.00 at 8-bit). This quantization aims to reduce memory footprint while preserving as much quality as possible for compatible inference engines supporting NVFP4.

Original author: FallenMerick
Original model link: FallenMerick/MN-Violet-Lotus-12B
License: CC-BY-4.0 (same as the original — please provide attribution when sharing or using derivatives)

Original Model Highlights (from FallenMerick)

This model excels at:

• Roleplaying with wide character personality adherence
• Creative and engaging outputs
• Noticeably high emotional intelligence
• Strong performance on local EQ benchmarks

It was created using mergekit with the Model Stock method (arxiv:2403.19522), blending several strong Mistral-Nemo-based merges/models anchored on mistralai/Mistral-Nemo-Instruct-2407.

Recommended Settings

Use the same settings recommended for the original model. Special thanks to @Zeldazachman for the excellent SillyTavern preset:

Quantization Details

• Method: NVFP4
• Base model quantized: FallenMerick/MN-Violet-Lotus-12B (full precision weights)
• Context length: 131072 tokens (inherited from original)
• Intended use: Local inference where memory is constrained but quality matters (e.g. roleplay / story focused apps)

For full details, merge recipe, benchmark images, and more — please visit the original repository:
→ https://huggingface.co/FallenMerick/MN-Violet-Lotus-12B

If you enjoy this quantized version, consider giving a like/star on the original repo as well — all credit goes to FallenMerick for the excellent merge!