Small open-weights multimodal model from OpenBMB for image and video understanding on phones and edge devices.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
MiniCPM-V 4.6 is a compact multimodal language model from OpenBMB, built on a SigLIP2-400M vision encoder and the Qwen3.5-0.8B language model. It accepts single images, multiple images and video together with text, and outputs text. It is released in Instruct and Thinking variants.
The model uses the LLaVA-UHD v4 technique, which OpenBMB says cuts visual encoding FLOPs by more than 50%. It supports mixed 4x/16x visual token compression, so users can trade accuracy for speed. According to the model card, it scores 13 on the Artificial Analysis Intelligence Index, ahead of Qwen3.5-0.8B (10) and Ministral 3 3B (11), and reaches Qwen3.5 2B-level results on several benchmarks such as OpenCompass, OCRBench and HallusionBench. These figures are vendor-reported.
It is designed for on-device use, with edge adaptation code open-sourced for iOS, Android and HarmonyOS. It works with vLLM, SGLang, llama.cpp and Ollama, and can be fine-tuned with SWIFT and LLaMA-Factory. Quantized variants are provided in GGUF, BNB, AWQ and GPTQ formats. Weights are on Hugging Face under Apache-2.0, and OpenBMB also offers a hosted API with a public free key, plus a demo Space.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Tiny vision-language model that runs on phones across iOS, Android and Harmony, with a 2.6 GB total size. It outperforms similar-sized models on document, chart, GUI and video understanding and is token-efficient.
Closest neighbours by meaning, not just by category.
Model · OpenBMB
1B-parameter dense open-weights language model from OpenBMB for on-device use, with hybrid thinking and 128K context.
Model · OpenBMB
Dense 2B-parameter open-weights language model from OpenBMB for on-device and edge deployment, Apache-2.0 licensed.
Model · MiniMax
Open-weights native multimodal MoE model from MiniMax, ~428B parameters (~23B active) with 1M-token context.
Model · Google
Google's open-weights 12B encoder-free multimodal model in the Gemma 4 family, taking text, image and audio input.
Model · Google
Google DeepMind's family of open-weights models built from Gemini technology, from phone-sized to laptop-class variants.
Model · Thinking Machines
Open-weights multimodal MoE model (276B total, 12B active) from Thinking Machines; text, image and audio in, text out.
