MiniCPM-V 4.5 successor ships first MoE variant: 35B-A3B runs 13+ tokens/s on iPhone with 3 GB
heyshrutimishra · x · 2026-10-16
OpenBMB released MiniCPM-V 4.7, a multimodal model in two variants: 35B-A3B (the series' first MoE model) and 1B, advancing the performance-efficiency Pareto frontier for multimodal models.
On-device highlights:
- On iPhone 17 Pro Max, 35B-A3B reports 13+ tokens/s using just 3 GB of memory
- On iPad Pro 11" M4 (16 GB), Lookahead Expert Scheduling boosts speed from 2.97 to 14.92 tokens/s
- The 1B's 4-bit GGUF is 672 MB; both run locally via Ollama
The author suggests developers building image-processing apps pick one repeated task, benchmark accuracy and speed on their actual hardware, and compare total cost against their current API. Code and quantized builds are open-sourced (26.8k GitHub stars).
Related event: OpenBMB Releases MiniCPM-V 4.7 with First MoE Model Running on Phones(4 posts)→
More from Infra
- Rapid-MLX 0.16.2 brings NVIDIA Nemotron 3.5 up to 120B to Apple Silicon Macs — kuchaev · 2026-10-16
- Adobe ships NVIDIA RTX Spark optimizations: up to 2x faster creative apps and 5x Frame.io gains — shashib · 2026-10-16
- Meta's commitments jumped to $349B in a quarter as server memory prices tripled ahead of Oct 21 earnings — tengyanAI · 2026-10-16
- Meta releases Muse Spark Open, trained with vLLM-powered RL inference from day one — colinmcnamara · 2026-10-16
- a16z: bot 403 blocks double as Amazon cracks down on agentic commerce — a16z Newsletter · 2026-10-16
- Meta's open Muse Spark needs 493GB of memory — far beyond any consumer laptop — TejasKumar_ · 2026-10-16