MiniCPM-V 4.7 hits HF trending: 35B-A3B MoE runs on phones in ~3GB
lmoroney · x · 2026-10-18
OpenBMB's MiniCPM-V 4.7 hit #3 on Hugging Face trending, offering a useful data point on how big a model a phone can actually run.
- The 35B-A3B MoE variant runs on iOS and Android in about 3GB of memory
- A 1B variant delivers strong OCR, math, and video understanding
The author explains that "A3B" means only 3B of the 35B parameters are active per token — sparse routing is why phone hardware can keep up. He suggests a project for mobile devs: point the 1B model at receipts or whiteboard photos and see how far fully offline inference gets.
Related event: OpenBMB's MiniCPM-V 4.7 Runs on Phones with ~3GB of Memory(2 posts)→
More from Infra
- vLLM and NVIDIA push Kimi K3 decode to 842 tokens/s on GB300 with fused CuTe DSL path — vllm_project · 2026-10-18
- SemiAnalysis's 1.68x vLLM speedup claim for Kimi K3 questioned over TP16 vs TP8 baseline — JiaZhihao · 2026-10-18
- Blackstone's AI empire: $200B data centers, $95B chip financing, bets on Anthropic and OpenAI — SuB8u · 2026-10-18
- Modeling tokens-to-value: solving agent harness configs via a single utility function — JoshPurtell · 2026-10-18
- Open-source Intro to Semiconductor Manufacturing book follows a wafer from sand to chip in 25 chapters — CatAstro_Piyush · 2026-10-18
- When every ablation you run costs $1M — gordic_aleksa · 2026-10-18