ModelBest describes MiniCPM-V 4.7 35B-A3B as the first open-source MoE multimodal model built to run smoothly on-device without compromising model quality. The model activates about 3 billion parameters per token and the most experts stay on SSD storage and load as needed, bringing large-model multimodal capability to IOS and Android. It handles general visual understanding, STEM problem solving, OCR, document and chart interpretation, and long-video understanding. According to ModelBest, the model pushes forward the intelligence-cost Pareto frontier for on-device multimodal AI by delivering strong overall performance with short inference time.

The 1B version builds on MiniCPM-V 4.6. Designed for lightweight edge devices, it supports general vision, OCR, chart reasoning and multi-image understanding. It can run smoothly on entry-level devices, bringing multimodal AI to more everyday hardware.

ModelBest's Lookahead Expert Scheduling (LES) predicts which experts the next layer will need and loads them while the current layer computes. In most scenarios, it prepares more than 80% of required experts in advance without additional training. The model's original router still makes the final expert selection, so outputs are identical with LES enabled or disabled.

LES works together with layer-wise quantization and contiguous expert storage. Combined, these optimizations increased average generation speed approximately 5x on an iPad Pro M4 with 16 GB of memory. The model reaches more than 17 tokens per second on that device and more than 13 tokens per second on the iPhone 17 Pro Max.

The model also supports both 4× and 16× visual token compression, allowing developers to prioritize detail or speed. At 16× compression, it retains strong performance, scoring 90.9 on MMBench and 90.4 on OCRBench.

A new Self-Replay Fine-Tuning method mixes the model's own high-quality responses into training data. This helps the model learn new tasks while retaining its general abilities, at a cost close to standard supervised fine-tuning.

MiniCPM-V 4.7 has Day-0 support from AMD, Apple, Arm, Houmo AI, Intel, Rockchip and NVIDIA. Through the FlagOS software stack, it also has Day-0 support for AI and edge chips from CIX Technology, Huawei, Hygon, Iluvatar CoreX, Kunlunxin, MetaX, Moore Threads, Sunrise, T-Head and TsingMicro. Model weights, inference code and fine-tuning code are publicly available for developers to download, evaluate and deploy.

Despite packing just 2 billion parameters, MiniCPM5-2B natively supports tool calling, deep search, code generation, and multi-step reasoning, pioneering the deployment of general-purpose agentic capabilities directly on edge devices. In evaluations by Artificial Analysis, MiniCPM5-2B clinched the #1 spot on the Intelligence Index among open-source models under 4 billion parameters. It also scored 20 on the Agentic Index—which gauges an AI's autonomous task-execution capability—significantly outpacing its peers in the sub-4B category.

This breakthrough spotlights a paradigm shift in AI development: rather than aggressively scaling parameter counts, boosting "intelligence density" allows resource-constrained devices to deliver powerhouse performance within minimal compute, memory, and power envelopes. This approach accelerates the transition of advanced AI from centralized clouds to everyday edge hardware.

The practical value of this architecture has already been proven across diverse real-world edge scenarios. Powered by MiniCPM5-2B, end-users and enterprises can seamlessly run document processing, data synthesis, code generation, and multi-turn complex Q&A locally on PCs, smartphones, robotics, and IoT hardware. Shifting compute to the edge not only safeguards user privacy and data sovereignty, but also dramatically slashes latency, cloud API costs, and infrastructure dependency.

Unlike conventional "weights-only" releases that function as black boxes, ModelBest and OpenBMB have fully open-sourced their end-to-end pipeline across data curation, pre-training, and alignment. This radical transparency equips global researchers and developers with a reproducible blueprint to fine-tune, build, and deploy specialized edge applications.

To date, downloads across ModelBest's MiniCPM family have surpassed 50 million. By sharing models, methodologies, frameworks, and raw recipes, ModelBest is deepening its commitment to the global open-source ecosystem—democratizing high-efficiency, lightweight, and deployable edge AI worldwide.

At the end of May this year, ModelBest, in collaboration with the OpenBMB open-source community, released ForgeTrain, a production-level pre-training framework engineered by AI with no human in the loop. Using ForgeTrain, the team trained the base model of MiniCPM5-1B, a leading small language model of the world according to AA Index, at a training speed reportedly 10% faster than NVIDIA's Megatron. Later on, ModelBest continued to announce two further RSI technologies as follows:

During the World Artificial Intelligence Conference (WAIC) in July, ModelBest developed an agent powered by GLM-5.2 to iteratively optimize the training framework in only 18 hours, which then trained an AI model named MiniCPM5-130M from scratch. The resulting model turned out to match Gemma 3 270M in performance.

In August, ModelBest released ForgeStencil, a fully automated stencil optimization and deployment system engineered by AI, also with no human in the loop. This AI system optimized more than 100 stencils for industrial and scientific software in just one week.

Behind these achievements is the new technical paradigm, Forge Engineering, proposed by ModelBest. For decades, general-purpose frameworks have been the dominant paradigm of software engineering, which relies on expensive human programmers and comes at the cost of performance. But now, with the development of AI coding, it is possible to forge customized software for every demand at a lower and lower price.

The dramatic boost in training efficiency and sharp drop in compute costs prove the huge potential of this new paradigm in the most direct way. This landscape could have far-reaching implications for the future of AI. By developing AI systems faster and cheaper, more and more users can benefit from advanced technologies.