Home
MingTang Intelligent Unveils MiniCPM-V 4.7 for Edge Devices, Delivering 13+ Token/s on 3GB Memory
Mingshi Intelligence unveiled its latest edge-side multimodal large model, MiniCPM-V 4.7, available in two configurations: 35B-A3B and 1B. The 35B-A3B variant marks the first Mixture of Experts (MoE) model in the MiniCPM-V series, leveraging a proprietary "forward-looking expert scheduling" technology to deliver lossless, smooth inference on edge devices, making it the first open-source MoE capable of running efficiently without cloud dependency.

The standout metric is its performance: on an iPhone 17 Pro Max, the model consumes only 3GB of memory while achieving a decoding speed exceeding 13 tokens per second. This efficiency enables local multimodal reasoning on high-end smartphones, eliminating the need to transmit data to the cloud and ensuring enhanced privacy and reduced latency.
Multi-platform support and open-source fine-tuning solutions
Regarding compatibility, version 4.7 supports iOS, Android, and HarmonyOS, and is optimized for chip architectures including Intel, AMD, Arm, Rockchip, and Huanmo, covering both mobile and local computing environments. Mingshi also released the Self-Replay Fine-Tuning (SRFT) solution, empowering developers to expand model capabilities at a lower cost while mitigating catastrophic forgetting, thereby streamlining secondary development.
Performance benchmarks are equally impressive: in Instruct mode, the 35B-A3B variant outperforms Qwen3.6 35B-A3B with faster inference times and higher overall scores, while the 1B version builds upon the previous 4.6 release, demonstrating significant advantages in overall scoring and inference efficiency over Qwen3.5 0.8B. As edge models become increasingly compact and competitive, Mingshi’s latest iteration further lowers the barrier to deploying MoE large models directly on smartphones.
Related article
Nvidia’s Vera chip: The US$200 billion gamble Jensen Huang wants you to notice
The Nvidia Vera chip rarely grabs headlines when earnings exceed expectations, yet it deserves the spotlight. With Q1 revenue hitting US$81.62 billion—surpassing analyst estimates of US$78.86 billion—and a Q2 guidance of US$91 billion, well above Wal
Gamma buys out Accel-backed Lica to boost design capabilities
Gamma, a leading presentation platform, has acquired Lica, an Accel-backed design startup, to establish a dedicated design research lab, according to exclusive reports from TechCrunch. Lica’s co-founders will spearhead this new initiative.Originally
Zhipu MaaS to Introduce Data Non-Preservation Mechanism, Content Destroyed After Use
On September 21, Zhipu’s MaaS platform announced the upcoming release of a "data content not stored" feature, providing enterprises and developers with enhanced data privacy safeguards. This mechanism ensures that input and output data generated duri
Related Special Topic Recommendations
Comments (0)
0/500
Mingshi Intelligence unveiled its latest edge-side multimodal large model, MiniCPM-V 4.7, available in two configurations: 35B-A3B and 1B. The 35B-A3B variant marks the first Mixture of Experts (MoE) model in the MiniCPM-V series, leveraging a proprietary "forward-looking expert scheduling" technology to deliver lossless, smooth inference on edge devices, making it the first open-source MoE capable of running efficiently without cloud dependency.

The standout metric is its performance: on an iPhone 17 Pro Max, the model consumes only 3GB of memory while achieving a decoding speed exceeding 13 tokens per second. This efficiency enables local multimodal reasoning on high-end smartphones, eliminating the need to transmit data to the cloud and ensuring enhanced privacy and reduced latency.
Multi-platform support and open-source fine-tuning solutions
Regarding compatibility, version 4.7 supports iOS, Android, and HarmonyOS, and is optimized for chip architectures including Intel, AMD, Arm, Rockchip, and Huanmo, covering both mobile and local computing environments. Mingshi also released the Self-Replay Fine-Tuning (SRFT) solution, empowering developers to expand model capabilities at a lower cost while mitigating catastrophic forgetting, thereby streamlining secondary development.
Performance benchmarks are equally impressive: in Instruct mode, the 35B-A3B variant outperforms Qwen3.6 35B-A3B with faster inference times and higher overall scores, while the 1B version builds upon the previous 4.6 release, demonstrating significant advantages in overall scoring and inference efficiency over Qwen3.5 0.8B. As edge models become increasingly compact and competitive, Mingshi’s latest iteration further lowers the barrier to deploying MoE large models directly on smartphones.
Nvidia’s Vera chip: The US$200 billion gamble Jensen Huang wants you to notice
The Nvidia Vera chip rarely grabs headlines when earnings exceed expectations, yet it deserves the spotlight. With Q1 revenue hitting US$81.62 billion—surpassing analyst estimates of US$78.86 billion—and a Q2 guidance of US$91 billion, well above Wal
Gamma buys out Accel-backed Lica to boost design capabilities
Gamma, a leading presentation platform, has acquired Lica, an Accel-backed design startup, to establish a dedicated design research lab, according to exclusive reports from TechCrunch. Lica’s co-founders will spearhead this new initiative.Originally
Zhipu MaaS to Introduce Data Non-Preservation Mechanism, Content Destroyed After Use
On September 21, Zhipu’s MaaS platform announced the upcoming release of a "data content not stored" feature, providing enterprises and developers with enhanced data privacy safeguards. This mechanism ensures that input and output data generated duri











