MiniCPM-V 4.7 runs a 35B multimodal model in 3GB of memory, fully on-device
alifcoder · x · 2026-10-17
MiniCPM-V 4.7 (35B total, 3B active) runs fully on-device in about 3GB of memory, reading real-world controls and images without any cloud calls. In one test, a photo of a basement humidifier panel let the model read the 43% LCD humidity (already within the 30–50% winter range), recommend dial settings of 4–5, and catch the easily-missed Blower Activation switch that must be ON. The author credits Lookahead Expert Scheduling. The Hugging Face collection ships 35B-A3B and 1B variants with GGUF, FP8, GPTQ, AWQ and int4 quantizations plus DSpark versions.
More from Infra
- MiniCPM-V 4.7 runs locally at 100 tok/s using only ~3 GB of memory — NielsRogge · 2026-10-17
- Power-management chips: the invisible layer deciding AI data center energy efficiency — Olivier__OG · 2026-10-17
- Full FP8 n-gram table visibly improves local Qwen3.8-Flash-Next at zero speed cost — sillynoobhorse · 2026-10-17
- Local Nemotron 3.5 on one A100 hits 169 tok/s, processes 100 clinical notes for 2.4 cents — MaziyarPanahi · 2026-10-17
- NVIDIA's open-source AIPerf benchmarking tool and Triton GPU tutorials for inference measurement — techNmak · 2026-10-17
- Official Triton GPU Programming Tutorials: Vector Add to Fused Attention — techNmak · 2026-10-17