Reportedly: ModelBest releases MiniCPM-V 4.7, an on-device multimodal model with 13+ token/s decoding at 3GB memory
Reported by IT Home · not yet confirmed by the company or a second independent outlet. We update this page when it is.
ModelBest released MiniCPM-V 4.7, an on-device multimodal model in 35B-A3B and 1B sizes. The 35B-A3B MoE version runs at 13+ token/s with 3GB memory on iPhone 17 Pro Max. It supports iOS, Android, HarmonyOS, and chips from Intel, AMD, Arm, Rockchip, and Houmo, IT Home reported.
China context
- Original name
- 面壁智能
- Outside China
- Open weights · ithome.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can access the open-source MiniCPM-V 4.7 weights and SRFT method to build on-device multimodal applications without relying on cloud APIs.
- For investors
- ModelBest's efficiency gains on mobile hardware may intensify competition in the on-device AI segment, affecting valuations of edge AI startups and chip vendors.
Translated from Chinese. Quotes and facts link to the original sources.
ModelBest announced MiniCPM-V 4.7, a new on-device multimodal large model. It comes in 35B-A3B and 1B variants; the 35B-A3B is the first MoE model in the MiniCPM-V series. The model uses LES (Lookahead Expert Scheduling) inference acceleration. The 35B-A3B version is claimed to be the first open-source MoE model supporting lossless on-device smooth operation, achieving 13+ token/s decoding speed with only 3GB memory on iPhone 17 Pro Max. MiniCPM-V 4.7 supports iOS, Android, and HarmonyOS, and is adapted to Intel, AMD, Arm, Rockchip, and Houmo chip platforms. The company also open-sourced the Self-Replay Fine-Tuning (SRFT) method to help developers extend model capabilities and reduce forgetting. In Instruct mode, the 35B-A3B version reportedly achieves a higher composite score than Qwen3.6 35B-A3B with shorter inference time; the 1B version improves over MiniCPM-V 4.6 and leads Qwen3.5 0.8B in composite score and inference efficiency.
The 35B-A3B model uses a Mixture-of-Experts architecture with LES (Lookahead Expert Scheduling) to reduce memory and latency. The reported 13+ token/s at 3GB memory on an iPhone 17 Pro Max suggests aggressive quantization or expert offloading. The open-sourced SRFT fine-tuning method targets low-cost capability extension and forgetting reduction.
On-device model developers can now run a 35B-parameter MoE multimodal model on a flagship phone with 3GB memory, lowering the hardware barrier for local AI features. This pressures rival open-weight models like Qwen3.6 35B-A3B and Qwen3.5 0.8B on efficiency and composite scores.
ModelBest positions MiniCPM-V 4.7 for mobile and edge deployment across iOS, Android, HarmonyOS, and multiple chip platforms. The open-source SRFT method may attract developers seeking low-cost customization. The model's efficiency claims could influence procurement for on-device AI applications.
Next signals to check: independent benchmarks of MiniCPM-V 4.7 against Qwen3.6 35B-A3B and Qwen3.5 0.8B, real-world latency and memory usage on non-Apple devices, and adoption of the SRFT method by developers. Also verify whether the claimed scores are reproducible outside ModelBest's own tests.