MiniCPM-V-4.6-GGUF

MiniCPM-V-4.6 is OpenBMB's most edge-deployment-friendly multimodal model to date, a pocket-sized MLLM built on SigLIP2-400M and the Qwen3.5-0.8B LLM, designed for ultra-efficient image and video understanding directly on phones. It inherits strong single-image, multi-image, and video understanding from the MiniCPM-V family while introducing mixed 4x/16x visual token compression and a LLaVA-UHD v4-based architecture that cuts visual encoding FLOPs by over 50%, delivering ~1.5x token throughput versus Qwen3.5-0.8B. Despite its compact size, it scores 13 on the Artificial Analysis Intelligence Index — beating Qwen3.5-0.8B (10) with 19x fewer tokens, Qwen3.5-0.8B-Thinking (11) with 43x fewer tokens, and even the larger Ministral 3 3B (11) — while reaching Qwen3.5-2B-level performance on benchmarks like OpenCompass, RefCOCO, HallusionBench, MUIRBench, and OCRBench. The model is deployable across iOS, Android, and HarmonyOS with fully open-sourced edge adaptation code, supports flexible slicing/frame-sampling parameters for balancing detail versus speed, integrates with inference frameworks including vLLM, SGLang, llama.cpp, and Ollama (with GGUF/BNB/AWQ/GPTQ quantized variants), supports fine-tuning via SWIFT and LLaMA-Factory, and is released under the Apache-2.0 license.

Model Files

llama.cpp