The 35B-A3B model scored 90.4 on OCRBench, exceeding Qwen3.6 35B-A3B by 5.2 points even while using 16x hybrid visual token compression. OpenBMB validated the series across 34 benchmarks for STEM reasoning and long-video comprehension, providing day-0 adaptations for Apple, Intel, AMD, Arm, and Rockchip hardware. The models are available for download and deployment on GitHub and Hugging Face.
Tuesday, October 20, 2026
Story
OpenBMB MiniCPM-V 4.7 Hits 17 Tokens Per Second on Mobile Hardware
AI8 days ago
The latest multimodal AI release from OpenBMB provides a 5x speed increase for local inference on high-end tablets. The MiniCPM-V 4.7 series, which consists of 35B-A3B and 1B models, utilizes a Lookahead Expert Scheduling technique to preload expert weights and reduce processing stalls. This architecture enables the 35B-A3B variant to reach 17 tokens per second on an iPad Pro M4 and 13 tokens per second on an iPhone 17 Pro Max while using only 3GB of memory.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.