MiniCPM-V 4.7 runs a 35B MoE multimodal model on phones with just 3GB memory
nevrekaraishwa2 · x · 2026-10-15
OpenBMB released MiniCPM-V 4.7, bringing strong multimodal AI fully on-device:
- The 35B-A3B MoE activates only 3B params per token and uses LES (Lookahead Expert Scheduling) to preload experts, running on phones with 3GB of memory — reading street signs, parsing messy documents, and reasoning over charts locally.
- A lightweight 1B variant targets resource-constrained hardware.
- SRFT (Self-Replay Fine-Tuning) lets developers customize for new tasks while preserving general performance; available on both iOS and Android.
More from Infra
- 7 System Design Algorithms to Know: Merkle Trees to CAP/PACELC — goyalshaliniuk · 2026-10-15
- Sipeed MaixStation product page goes live for edge AI dev — xlltt · 2026-10-15
- Nvidia Is Now Worth More Than the Entire Indian Stock Market — theteknosaur · 2026-10-15
- Apple Silicon GEMM benchmarks: M6 does fp8 at ~2x fp16, M5 doesn't — Brilliant-Hall1387 · 2026-10-15
- NVIDIA joins EU's AI in Science Enablers, offering compute, training and cloud resources — nordicinst · 2026-10-15
- Open source's hardware trap: data centers crowd out consumer GPUs, local LLMs become a privilege — Dante_77A · 2026-10-15