MiniCPM-V 4.7: A 35B Multimodal MoE Running on Phones With ~3GB Memory
Equivalent-Grass-527 · reddit · 2026-10-14
OpenBMB/ModelBest released MiniCPM-V 4.7, an on-device multimodal push: a 35B-A3B MoE vision-language model that runs smoothly on phones and tablets with 3GB memory, with day-one support across multiple chip platforms.
- LES (Lookahead Expert Scheduling): instead of just shrinking the model, it predicts and preloads MoE experts to remove inference bottlenecks.
- Demoed skills: reading real-world signs, understanding complex documents, visual reasoning, and planning under real-world constraints.
- A lighter MiniCPM-V 4.7 1B targets more constrained devices.
- New Self-Replay Fine-Tuning (SRFT) lets developers customize for specific tasks while preserving general capabilities.
Weights and code are open on GitHub/Hugging Face — capable multimodal assistants moving off the cloud and onto the phone.
Related event: OpenBMB Open-Sources MiniCPM-V 4.7: 35B MoE Multimodal on Device(22 posts)→
More from Infra
- 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
- CoreWeave plunges 20% as credit markets price >60% default odds; is this dot-com all over again? — kimmonismus · 2026-10-16