OpenBMB releases MiniCPM-V-4.7-1B, a 1B-parameter multimodal model for on-device use
OpenBMB released MiniCPM-V-4.7-1B, a 1B-parameter multimodal model, on Hugging Face under Apache-2.0. It supports image-text-to-text tasks and runs on iOS, Android, and HarmonyOS. The release includes GGUF, BNB, AWQ, GPTQ, and FP8 quantized variants.
China context
- Original name
- 面壁智能
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can integrate a 1B-parameter multimodal model with Apache-2.0 license and support for iOS, Android, HarmonyOS, and multiple chips, reducing deployment friction for on-device AI.
- For investors
- The release of a small, efficient multimodal model with broad platform support may pressure competitors in the on-device AI market and signal ModelBest's focus on edge deployment.
OpenBMB released MiniCPM-V-4.7-1B, a 1B-parameter multimodal model, on Hugging Face under Apache-2.0. The model supports image-text-to-text tasks and is designed for on-device inference, with support for iOS, Android, and HarmonyOS. It includes quantized variants (GGUF, BNB, AWQ, GPTQ, FP8) and is compatible with frameworks such as SGLang, vLLM, llama.cpp, and Ollama.
MiniCPM-V-4.7-1B uses intra-ViT early compression from LLaVA-UHD v4 to reduce visual token computation, and supports mixed 4x/16x visual token compression. The model card claims it outperforms Qwen3.5 0.8B in overall score and inference efficiency in Instruct mode, but this is company-reported and not independently verified.
Developers outside China can now use a 1B-parameter multimodal model with Apache-2.0 licensing and broad hardware support, reducing the cost and complexity of deploying vision-language models on edge devices. This directly competes with other small open-weights models like Qwen3.5 0.8B for on-device AI applications.
The model's small size, permissive license, and support for multiple deployment frameworks and chips make it attractive for companies building on-device AI features without incurring high inference costs or vendor lock-in.
Watch for independent benchmarks comparing MiniCPM-V-4.7-1B against Qwen3.5 0.8B and other small multimodal models, and for adoption in on-device applications on Rockchip, Intel, AMD, and Arm platforms.