How MiniCPM-V 4.6 Achieves Lightning‑Fast Multimodal AI on Smartphones (Open‑Source)

MiniCPM‑V 4.6 is an open‑source multimodal model that combines a SigLIP2‑400M visual encoder with a Qwen3.5‑0.8B large language model. The development team deeply re‑engineered the internal computation flow to cut compute demand.

Compute‑efficiency breakthrough

On the Artificial Analysis Intelligence Index benchmark the model scores 13 points, higher than the same‑size Qwen3.5‑0.8B (10) and its Thinking variant (11). Compared with those baselines, MiniCPM‑V 4.6 reduces token‑processing cost by 19× and 43× respectively and also outperforms the larger 3 B Ministral model (score 11). The improvement stems from adopting the latest LLaVA‑UHD v4 techniques, which lower visual‑encoding FLOPs by more than 50 %. The reduced arithmetic load translates into higher energy efficiency and a token‑throughput increase of roughly 1.5× over Qwen3.5‑0.8B. First‑token latency on a 3136×3136 image is 75 ms.

Multimodal capabilities

The model retains the series’ strength in single‑image analysis, multi‑image fusion, and dynamic video understanding. Across OpenCompass, RefCOCO, HallusionBench, MUIRBench and OCRBench the model consistently surpasses Qwen3.5‑0.8B and reaches the capability level of a 2 B‑scale model. To balance speed and visual fidelity, MiniCPM‑V 4.6 introduces a mixed visual‑token compression ratio ranging from 4× to 16×, allowing runtime selection between higher quality analysis and faster response.

Edge deployment

MiniCPM‑V 4.6 is fully compatible with iOS, Android and HarmonyOS. All adaptation code is released under an open‑source license, enabling developers to follow the provided guide and reproduce smooth AI interactions on phones or tablets.

Ecosystem integration and quantization

Inference support includes vLLM, SGLang, llama.cpp and Ollama. Fine‑tuning ecosystems such as SWIFT and LLaMA‑Factory are also compatible. To accommodate diverse hardware, the release provides quantized model files in GGUF, BNB, AWQ and GPTQ formats.

Model hub: https://huggingface.co/openbmb/MiniCPM-V-4.6

Source code: https://github.com/OpenBMB/MiniCPM-V

Mobile‑app repository: https://github.com/OpenBMB/MiniCPM-V-Apps