MiniCPM-V 4.6 is a 1.3B-parameter model supporting 4x and 16x visual token compression. Against Qwen3.5-0.8B on standard image input, visual encoding FLOPs drop by over 50% and token throughput is about 1.5x that of Qwen3.5-0.8B.
Highlights
Base and multimodal benchmark results：It scores 13 on the Artificial Analysis Intelligence Index 4.0, outperforming Qwen3.5-0.8B (10 points, 19× lower token cost), Qwen3.5-0.8B-Thinking (11 points, 43× lower cost), and Ministral 3 3B scores 11. Across OpenCompass, RefCOCO, HallusionBench, MUIRBench, and OCRBench, it reaches Qwen3.5 2B-level multimodal capability.
On-device architecture：Built on LLaVA-UHD v4,, it cuts visual-encoding FLOPs by more than 50% and delivers roughly 1.5× the token throughput of Qwen3.5-0.8B. Switchable 4×/16× visual-token compression lets developers balance accuracy and speed on edge devices.
Cross-Platform and Developer Friendly：Deploy to iOS, Android, and HarmonyOS with open-source edge-adaptation code. It supports vLLM, SGLang, llama.cpp, and Ollama, integrates with SWIFT and LLaMA-Factory fine-tuning ecosystems, and provides GGUF, BNB, AWQ, and GPTQ quantized variants.
Technical highlights
Hybrid 4×/16× Visual Token Compression：Switchable hybrid visual-token compression supports 4× for high fidelity and 16× for high efficiency. Developers can balance recognition accuracy and inference speed by scenario. Single images, multiple images, and video share a unified encoding, keeping capability and knowledge consistent across the three input forms.
LLaVA-UHD v4 vision architecture：Built on LLaVA-UHD v4, it reduces visual encoding FLOPs by over 50% and brings the 1.3B model's token throughput to about 1.5x that of Qwen3.5-0.8B.
Strong-Foundation Multimodal Capability：Built on SigLIP2-400M and Qwen3.5-0.8B, it inherits MiniCPM-V single-image, multi-image, and video understanding. It scores 13 on the Artificial Analysis Intelligence Index and reaches Qwen3.5 2B-level performance on OpenCompass, OCRBench, HallusionBench, and MUIRBench.
Benchmarks
General Capabilities
STEM Reasoning
Documents
Visual Grounding
Multi-Image Understanding
Hallucination
GUI
Video Understanding
Inference Efficiency · Throughput
* Accuracy is reported from official reports or local reproductions using official weights, evaluated in instruct mode. Inference efficiency is measured with vLLM on an NVIDIA RTX 4090; MiniCPM-V 4.6 uses 16× compression by default.