base_model:
- openbmb/MiniCPM-V-4.7-1B library_name: gguf pipeline_tag: image-text-to-text tags:
- vision
- multimodal
- minicpm-v
- abliterated
- heretic
- uncensored
- gguf
- llama.cpp
- ollama license: apache-2.0 language:
- en
- zh
MiniCPM-V-4.7-1B-abliterated-heretic-GGUF
This repository provides quantized GGUF checkpoints for MiniCPM-V 4.7 1B, based on the refusal-abliterated weights created via Heretic v1.4.0.
MiniCPM-V 4.7 1B is an ultra-efficient on-device multimodal large language model (MLLM) developed by OpenBMB. This version has been neutralized against standard refusal directions while retaining full multimodal vision and reasoning capabilities.
📊 Benchmark & Safety Metrics
The ablation was performed using Heretic v1.4.0 to eliminate refusal bias while keeping Kullback–Leibler (KL) divergence near reference baseline:
| Metric | Abliterated Model | Original Base Model |
|---|---|---|
| Refusal Rate | 9 / 100 | 99 / 100 |
| KL Divergence | 0.0191 | 0.0000 (Reference) |
A refusal rate reduction from 99% to 9% with an ultra-low KL divergence of 0.0191 ensures that the model complies with benign, complex, and unconstrained instructions without degrading language coherence or visual understanding.
📦 Quantization Files & Details
Important: To run multimodal image tasks, you must download
mmproj-model-f16.ggufalongside whichever text quant you select below.
| File | Size | Description & Recommended Use |
|---|---|---|
MiniCPM-V-4.7-1B-abliterated-heretic-Q8_0.gguf | 1.08 GB | Near-lossless quality. Best choice for systems with standard RAM. |
MiniCPM-V-4.7-1B-abliterated-heretic-Q6_K.gguf | 838 MB | Excellent balance of speed and high reasoning fidelity. |
MiniCPM-V-4.7-1B-abliterated-heretic-Q5_K_M.gguf | 752 MB | Great quality retention with reduced memory overhead. |
MiniCPM-V-4.7-1B-abliterated-heretic-Q4_K_M.gguf | 672 MB | Recommended default for on-device and mobile edge deployment. |
MiniCPM-V-4.7-1B-abliterated-heretic-Q3_K_M.gguf | 575 MB | Ultra-compact quant for memory-constrained environments. |
mmproj-model-f16.gguf | 1.11 GB | Vision projector weights (Required for all visual inference). |
🤝 Credits & Acknowledgements
- Original Base Model: OpenBMB for the release of MiniCPM-V-4.7-1B.
- Abliteration Framework: Heretic v1.4.0.