MiniCPM-V 4.7 runs a 35B MoE multimodal model on phones with 3GB of memory
alifcoder · x · 2026-10-14
OpenBMB released MiniCPM-V 4.7, bringing a 35B-A3B MoE multimodal model to on-device inference with just 3GB of memory on iOS and Android.
Key details:
- LES (Lookahead Expert Scheduling) is the inference engineering that makes 35B-class MoE efficient on mobile
- Live demos cover real-world sign reading, constraint-aware planning, complex document parsing and visual reasoning
- A 1B variant targets resource-constrained hardware
- Self-Replay Fine-Tuning (SRFT) lets developers customize for new tasks while preserving general performance
The resharing author notes the expert-scheduling inference work is the real highlight, not just the model size.
Related event: ModelBest Open-Sources MiniCPM-V 4.7: 35B MoE Multimodal Runs on Phones(18 posts)→
More from Infra
- Liquid AI ships LFM2.5-24B-A2B: 2B active params, 131K context, local agent model — maximelabonne · 2026-10-14
- $12,000 Dual Intel Arc B70 Pro Rig Runs GLM-5.3-Flash at ~100 tok/s — burny_tech · 2026-10-14
- Nvidia-backed Lambda raising $4B pre-IPO round at $14.5B valuation — Beth_Kindig · 2026-10-14
- IBM's STAIR cuts hallucination 65x by having the LLM read the table of contents — alex_verem · 2026-10-14
- Atomic Machines exits stealth with $250M raised and an AI-native micro-manufacturing platform — jfiance · 2026-10-14
- DigUp: open-source Mac app brings 100% local semantic search via EmbeddingGemma — TheMoonMidas · 2026-10-14