maternion/MiniCPM-V4.7-1B-Heretic-GGUF

MiniCPM-V 4.7 1B Heretic GGUF

GGUF quantizations of an abliterated MiniCPM-V 4.7 1B. Refusal directions were identified and surgically removed from the model weights using heretic's direction-ablation pipeline, reducing refusal behavior while keeping the base model's capabilities intact. Ready to run with llama.cpp, Ollama, LM Studio, and any other GGUF-compatible runtime.

Works with the base model's vision projector (mmproj), so image input is fully supported.

Heretic Details

Property	Value
Base model	openbmb/MiniCPM-V-4.7-1B
Method	Direction ablation (heretic)
Refusal prompts	LLM-LAT/harmful-dataset
Good prompts	mlabonne/harmless_alpaca
Trials	404 (TPE, 60 random startup)
Refusals	21/100
KL divergence	0.0122

The low KL divergence (0.0122) indicates the abliterated model stays numerically close to the base, preserving reasoning, tool-calling, and vision capabilities while reducing refusals.

Note: No model is fully "uncensored." A small residual refusal rate remains on the most deeply-trained refusal categories — this is expected and preferable to the capability loss incurred by more aggressive ablation.

Available Files

File	Quant	Size
MiniCPM-V-4.7-1B-Heretic.Q4_K_M.gguf	Q4_K_M	641 MB
MiniCPM-V-4.7-1B-Heretic.Q6_K.gguf	Q6_K	800 MB
MiniCPM-V-4.7-1B-Heretic.Q8_0.gguf	Q8_0	1.1 GB
MiniCPM-V-4.7-1B-Heretic.F16.gguf	F16	1.9 GB
mmproj-F16.gguf	vision projector	1.1 GB

Usage

llama.cpp

llama-server -m MiniCPM-V-4.7-1B-Heretic.Q8_0.gguf \
  --mmproj mmproj-F16.gguf \
  -c 32768 -ngl 99 --jinja --host 127.0.0.1 --port 8080

Ollama

The same quants are also published on the Ollama registry:

ollama run maternion/minicpm-v4.7-heretic:1b

Tag	Precision
1b	Q4_K_M
1b-q4_K_M	Q4_K_M
1b-q6_K	Q6_K
1b-q8_0	Q8_0
1b-f16	F16

---

MiniCPM-V 4.7

GitHub · Evaluation · Examples

MiniCPM-V 4.7 is the latest model in the MiniCPM-V series, available in a 1B variant. It builds on the family's single-image, multi-image, and video understanding capabilities, combining stronger foundation capabilities with efficient visual encoding and on-device inference. Notable features of MiniCPM-V 4.7 include:

• 🔥 A New Pareto Frontier in Performance and Efficiency. The 1B variant of MiniCPM-V 4.7 advances the performance–efficiency Pareto frontier for MLLMs. In Instruct mode, the 1B variant continues to improve foundation capabilities over MiniCPM-V 4.6 and substantially outperforms Qwen3.5 0.8B in both overall score and inference efficiency.
• 🚀 Efficient Architecture. Based on the intra-ViT early compression technique from LLaVA-UHD v4, MiniCPM-V 4.7 compresses visual tokens early within the vision encoder, reducing computation in subsequent encoder layers. It also supports mixed 4x/16x visual token compression.
• 💻 Broad Platform and Chip Support. MiniCPM-V 4.7 supports iOS, Android, and HarmonyOS, together with open-source on-device deployment support, and day-0 support for Rockchip, Houmo, Intel, AMD, and Arm chips.

Evaluation

Significantly Advancing the Performance–Efficiency Pareto Frontier

The 1B variant of MiniCPM-V 4.7 advances the performance–efficiency Pareto frontier. In Instruct mode, 1B substantially outperforms Qwen3.5 0.8B in both overall score and inference efficiency.

Overall Performance (1B, Instruct)

Examples

MiniCPM-V 4.7 can run on phones and tablets. The clip below is an unedited screen recording captured directly on the device.

1B iPhone 17 Pro Max

Acknowledgements

We would like to thank the following projects:

• Qwen3.5 and Qwen3.6 for providing the language backbone
• SigLIP2 for providing the vision understanding module
• Transformers