MiniCPM-V 4.7

MiniCPM-V 4.7 comes in two versions: 1B and 35B-A3B. The 35B MoE runs smoothly on iOS and Android with only about 3 GB of memory, while the 1B version ranks in the top tier of its size class for OCR, math, and video understanding.

Highlights

• SOTA MiniCPM-V 4.7 — 35B MoE on Your Phone：With Lookahead Expert Scheduling (LES) and related techniques, it runs smoothly on iPhone / iPad and flagship Android phones with only about 3 GB of memory.
• All-Round Vision, Raising the On-Device Ceiling：It matches flagship models of the same tier and even larger ones on major benchmarks in general vision, STEM, OCR & charts, long video, and Chinese knowledge. Even better, scores barely drop with 16× visual token compression (MMBench still 90.9, OCRBench still 90.4).
• 1B: Small but Mighty：OCRBench 86.1, MathVista 74.3, CharXiv Reasoning 48.8, and for video understanding, LongVideoBench 53.4 and MotionBench 54.55. Against Qwen3.5-0.8B and even the larger LFM2.5-VL 1.6B/3B, it leads across MMStar, MathVista, OCR, chart reasoning, hallucination suppression, and video understanding — an ideal base for vertical-domain adaptation and lightweight on-device deployment.

Technical highlights

• Lookahead Expert Scheduling (LES)：Running an MoE on device is only the starting point — the real challenge is preserving model capability and smooth generation speed at the same time. Storing expert weights on SSD or other storage breaks through memory limits, but if loading and computation can't run in parallel, frequent waits for loading slow generation down; capping the number of experts or fixing the expert set speeds things up, but can degrade capability. MiniCPM-V 4.7 35B-A3B uses LES to jointly optimize expert loading, compute scheduling, and weight access, and with DSpark speculative-decoding support it achieves lossless inference at 13+ tokens/s on iPhone 17 Pro Max with only 3 GB of memory.
• Efficient Visual Encoding from the Series：Architecturally, MiniCPM-V 4.7 carries over two proven techniques from the series: slice-based image encoding (high-resolution images are cut into 448×448 slices plus a low-resolution overview, so attention cost grows linearly with the number of slices) and mixed 4×/16× visual token compression (4× keeps more detail, 16× infers faster, switchable per task).

Benchmarks

• 35B-A3B · General：MiniCPM-V 4.7 35B-A3B leads almost across the board in general vision (MMBench 91.7, MMStar 80.7), STEM (MMMU 80.2, MathVista 88, MathVerse 86.5), OCR & charts (OCRBench 91.2, CharXiv Reasoning 75.6), long video (LongVideoBench 68.9), and Chinese knowledge (CMMLU 90.2). With 16× visual compression it still scores MMBench 90.9 and OCRBench 90.4, a minimal loss.
• 35B-A3B · STEM
• 35B-A3B · Document, OCR & Chart
• 35B-A3B · Multi-Image & Video
• 35B-A3B · Hallucination
• 35B-A3B · GUI
• 35B-A3B · Instruction Following
• 1B · General：The 1B version is a comprehensive upgrade over MiniCPM-V 4.6-1.3B and surpasses Qwen3.5-0.8B and LFM2.5-VL-3B, staying ahead of its size class in OCR, chart reasoning, and video understanding.
• 1B · STEM
• 1B · Document, OCR & Chart
• 1B · Multi-Image & Video
• 1B · Hallucination
• 1B · GUI
• 1B · Instruction Following

* 4× / 16×: visual token compression ratio. The higher the ratio, the fewer visual tokens an image is encoded into. MiniCPM-V 4.7 (4×) / (16×) in the charts are its results at each ratio. * MiniCPM-V 4.7 support mixed 4×/16× visual token compression, with either ratio selectable; other models support only fixed compression ratios.