OpenBMB has released MiniCPM-V 4.7, a new vision-language model series that marks the debut of a sparse mixture-of-experts architecture within the company’s mobile-focused lineup. The headline release is the 35B-A3B variant, a model designed to balance high-level visual understanding with the strict power and memory constraints of modern smartphones.
Navigating the MiniCPM-V 4.7 Series
For developers and researchers attempting to track down the model, the naming conventions have caused some confusion. According to reporting by OrcaRouter, the name “MiniCPM-V 4.7” serves as a banner for a broader collection of repositories rather than a single downloadable file. A direct search for a repository named “openbmb/MiniCPM-V-4.7” returns an authorization error, as no such bare checkpoint exists. Instead, the series includes thirteen different repositories, ranging from the base 35B-A3B checkpoint to various quantized versions like GGUF, AWQ, and GPTQ.
The specific model generating interest is the 35B-A3B checkpoint. OpenBMB created this repository on October 7, 2026, and later added a comprehensive model card on October 12, 2026. This model is released under the Apache-2.0 license, which makes it a viable option for developers building commercial applications. This stands in contrast to other recent releases in the field, such as NVIDIA’s PixelUMM, which ships under a more restrictive non-commercial license.
How the Mixture-of-Experts Architecture Works
The 35B-A3B model is the first in the MiniCPM-V series to utilize a sparse mixture-of-experts, or MoE, architecture. In plain terms, think of an MoE model like a specialized office building where only a few experts are called into a meeting room at any given time, rather than requiring the entire staff to be present for every single task. This allows the model to contain a large number of parameters while only activating a small, efficient subset for each individual token generated.
To keep this process smooth on mobile hardware, OpenBMB introduced a technique called Lookahead Expert Scheduling. According to OpenBMB, this method works alongside fine-grained layer-wise quantization and DSpark speculative sampling to prevent the bottlenecks that usually plague large models on handheld devices. The practical upshot is that the model can maintain a decoding speed of over 13 tokens per second while consuming only 3 GB of memory on an iPhone 17 Pro Max.
The model also employs an efficient visual processing pipeline. It uses intra-ViT early compression, a technique credited to LLaVA-UHD v4, to shrink visual tokens within the vision encoder before they reach the main language stack. This reduces the computational load on subsequent layers. The shared vision tower consists of 27 layers with a hidden size of 1,152 and a patch size of 14, operating at an image size of 980. Developers can further manage the balance between visual detail and inference speed by choosing between 4x or 16x visual token compression settings.
Platform Compatibility and Developer Tools
OpenBMB has prioritized broad hardware support for this release. The model is compatible with iOS, Android, and HarmonyOS, with day-zero support for a variety of chip platforms including Rockchip, Houmo, Intel, AMD, and Arm. This wide compatibility is intended to extend the reach of high-performance visual understanding to a variety of local computing devices beyond traditional desktop workstations.
For those looking to integrate or customize the model, the developer ecosystem is extensive. The model supports standard inference frameworks such as SGLang, vLLM, llama.cpp, and Ollama. For users interested in fine-tuning, the company provides support through SWIFT and LLaMA-Factory. OpenBMB has also introduced a method called Self-Replay Fine-Tuning, which is designed to help developers add new capabilities to the model while minimizing the risk of “catastrophic forgetting,” where the model loses its previous knowledge during the training process.
Positioning in the Current Market
The release of MiniCPM-V 4.7 arrives during a period of rapid development for on-device artificial intelligence. According to OpenBMB, the 35B-A3B variant establishes a new Pareto frontier for performance and efficiency in multi-modal large language models. In internal testing, the model achieved a higher overall score in Instruct mode compared to the Qwen3.6 35B-A3B model, while requiring less time for inference.
Because the field is moving quickly, it is worth noting that independent, third-party benchmarks for these specific claims are not yet widely available. While the model card provides detailed performance metrics, the current landscape of vision-language models remains highly competitive. For developers, the choice between models like MiniCPM-V 4.7 and other emerging architectures often comes down to the specific licensing requirements and the hardware targets of their projects. With its Apache-2.0 license and focus on mobile-native optimization, MiniCPM-V 4.7 is positioned as a tool for those who need to deploy high-capability visual understanding directly onto user devices.
Understanding the MiniCPM-V 4.7 Lineage
The MiniCPM-V 4.7 series encompasses more than just the 35B-A3B variant. According to documentation, the series includes a second, smaller dense model, creating a family that shares a lineage name and a common vision tower while differing in their underlying architecture. While the 35B-A3B is a sparse mixture-of-experts, this smaller counterpart provides an alternative for different resource profiles. Both models utilize the same minicpmv4_7_vision tower, which features 27 layers, a hidden size of 1,152, and a patch size of 14, operating at an image size of 980.
The release timeline for these assets is marked by a distinction between repository creation and documentation availability. While the repositories were established on October 6 and 7, 2026, the model card, license tags, and performance charts were not added until October 12, 2026. This five-to-six-day gap between the initial weight uploads and the finalized documentation is a critical detail for those tracking the project’s history. The series also includes support for FP8 quantization, expanding the range of available formats beyond the initial GGUF, AWQ, and GPTQ variants.
Sources
- openbmb/MiniCPM-V-4.7-35B-A3B · Hugging Face
- MiniCPM-V 4.7 35B-A3B vs MiniCPM-V 4.7: Which Exists?
- MiniCPM-V 4.7 35B-A3B vs PixelUMM: The Licence First
Photo: Ahmad Ali Karim (CC0)
