openbmb released MiniCPM-V-4.7-35B-A3B, the first MoE model in the MiniCPM-V series
openbmb released MiniCPM-V-4.7-35B-A3B, the first MoE model in the MiniCPM-V series, on Hugging Face. It achieves 13+ tokens/s decoding with 3 GB memory on iPhone 17 Pro Max. 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 access the model weights on Hugging Face and deploy on iOS, Android, HarmonyOS, and supported chips, with quantized variants and multiple inference frameworks.
- For investors
- The release demonstrates ModelBest's capability in efficient on-device MoE, potentially attracting partnerships with device manufacturers and edge AI platform providers.
openbmb released MiniCPM-V-4.7-35B-A3B, the first MoE model in the MiniCPM-V series, on Hugging Face. The model supports single-image, multi-image, and video understanding, and introduces Lookahead Expert Scheduling (LES), fine-grained layer-wise quantization, and DSpark speculative sampling integration to enable smooth on-device inference. It achieves 13+ tokens/s decoding with 3 GB memory on iPhone 17 Pro Max, and supports iOS, Android, HarmonyOS, and chip platforms from Rockchip, Houmo, Intel, AMD, and Arm. The release includes GGUF, BNB, AWQ, GPTQ, and FP8 quantized variants, and supports inference frameworks such as SGLang, vLLM, llama.cpp, and Ollama.
MiniCPM-V-4.7-35B-A3B uses Lookahead Expert Scheduling (LES) to prefetch experts based on cross-layer routing prediction, overlapping computation with loading of remaining experts. It also employs fine-grained layer-wise quantization and DSpark speculative sampling to reduce memory traffic and improve decoding efficiency. The model supports mixed 4x/16x visual token compression via intra-ViT early compression from LLaVA-UHD v4.
Developers outside China can now deploy a 35B-A3B MoE multimodal model on-device with 3 GB memory, reducing cloud inference costs for mobile applications. This pressures other open-weights labs to match on-device MoE efficiency or lose edge deployment opportunities.
The model enables low-cost on-device multimodal AI for mobile and edge applications, potentially reducing reliance on cloud APIs and enabling new use cases in privacy-sensitive or offline environments. The broad platform and chip support lowers integration barriers for device manufacturers and app developers.
Watch for independent benchmarks comparing MiniCPM-V-4.7-35B-A3B against Qwen3.6 35B-A3B in Instruct mode, and for adoption of the LES technique in other MoE models. Also monitor whether the claimed on-device performance holds across different hardware beyond iPhone 17 Pro Max.