MiniCPM-V 4.6's 1.3B Model Dimension Reduction Attack Redefines Edge Multimodal Technology

On May 11, Mingshi Intelligence teamed up with Tsinghua University and the OpenBMB open-source community to unveil the new edge-side multimodal large model MiniCPM-V4.6. Despite its compact size—just 1.3 billion parameters—this lightweight model pushes the boundaries of larger models through exceptional intelligent density and cross-platform adaptability, speeding up the real-world deployment of edge-side AI.

1. Peak Performance: 1.3B Parameters Deliver Exceptional Results

MiniCPM-V4.6 comes in two versions—Instruct and Thinking—both demonstrating impressive reasoning and understanding across various benchmarks when compared to similarly sized models.

Global Leadership: On the Artificial Analysis (AA) leaderboard, MiniCPM-V4.6 scored an impressive 13 points, far surpassing similarly sized rivals (e.g., Alibaba's Qwen3.5-0.8B and Google's Gemma4-E2B-it) while nearly matching the performance of larger models like Qwen3.5-2B. This sets a new benchmark for 1B-parameter models.

Advanced Capabilities: From general image-text comprehension and complex STEM math reasoning to demanding document OCR and video temporal understanding, the model exhibits strong intelligence. The Thinking version, in particular, excels in multi-image reasoning and hallucination suppression.

2. Efficiency Breakthrough: Extreme Intelligent Density at the Edge

To tackle memory constraints in edge deployment, MiniCPM-V4.6 has been deeply optimized for inference speed and resource efficiency.

Low Memory Footprint: Memory requirements drop to just 6GB, enabling smooth operation on mainstream smartphones, PCs, and smart home devices.

Inference Efficiency: Powered by vLLM, inference throughput is 1.5× that of competitors. When processing a 3136² ultra-high-definition image at the edge, the first response latency is just 75.7ms—2.2× faster than competing models.

Throughput: A single GPU delivers text generation at 7,013 tokens per second and can process 1,344² images at 54.79 images per second, demonstrating remarkable efficiency.

3. Core Technology: LLaVA-UHD v4 Minimizes Overhead

The model's lightweight design is made possible by LLaVA-UHD v4, jointly developed by Mingshi Intelligence and Tsinghua University.

Encoding Redesign: Through a revamped ViT image encoding and shallow compression module, image encoding overhead is cut by 50% and high-resolution floating-point operations by 55.8%.

Hybrid Compression: It introduces a 4×/16× hybrid token compression that enables flexible switching between performance-first and speed-first modes. This technology has been validated in Kuaishou's recommendation model OneRec, handling massive traffic.

4. Ecosystem Deployment: From Lab to Industry

The open-source release of MiniCPM-V4.6 marks both a technical and ecosystem milestone.

Easy Development: It integrates seamlessly with fine-tuning frameworks like ms-swift and LLaMA-Factory, enabling full-scale fine-tuning on a single RTX 4090 GPU.

Cross-Platform Support: Compatible with vLLM, Ollama, and other major frameworks, it offers test versions for iOS, Android, and HarmonyOS, bringing AI to a wider range of hardware.

Real-World Impact: The model series is already deployed in automotive, PC, smart home, and industrial inspection sectors, with partners including Lenovo, Geely, SAIC Volkswagen, Xiaomi, and OPPO.