OpenBMB launches MiniCPM-V 4.7: first open-source MoE MLLM series with smooth on-device inference (35B-A3B / 1B), ~13+ tok/s on iPhone 17 Pro Max

Why it matters

【Key points】

OpenBMB on Oct 12 announced MiniCPM-V 4.7, an open multimodal series for on-device use. Two sizes: 35B-A3B (first MoE in the line, ~3B active) and a light 1B. HF weights appeared Oct 7; the official X post landed at 17:15Z Oct 12.

Headline technique: Lookahead Expert Scheduling (LES) prefetches experts across layers, with layer-wise quantization and DSpark speculative sampling, cutting load stalls while keeping original token-level routing. Claims: ~5× speedup on iPad Pro M4 16GB with byte-identical outputs (17+ tok/s); 13+ tok/s on iPhone 17 Pro Max at ~3GB. Hybrid 4×/16× visual compression; at 16× they report OCRBench 90.4 vs Qwen3.6 35B-A3B. Apache-2.0; iOS/Android/HarmonyOS; Day-0 for Rockchip/Intel/AMD/Arm; SGLang, vLLM, llama.cpp, Ollama.

【Verifiable】

@OpenBMB (2026-10-12 17:15Z, ~76k views / 729 likes): https://x.com/OpenBMB/status/2109694573408866351

HF collection: https://huggingface.co/collections/openbmb/minicpm-v-47

Repo openbmb/MiniCPM-V-4.7-35B-A3B (created 2026-10-07, Apache-2.0)

Docs: https://github.com/OpenBMB/MiniCPM-o/blob/main/docs/minicpm_v4dot7_en.md

Narrow X heat ~79 posts in ~5h, rising; not previously on agipulse.org.

【Our read】

1. The product constraint is smooth MoE under phone memory/bandwidth—LES + quantization + speculative sampling, not shrinking the expert pool.

2. Weights preceded the X narrative; easy to miss if patrols only watch fresh HF repos. Care about reproducible tok/s and open stacks, not news-card hype.

3. Same cycle as Muse Wonder; next filter is community battery/hallucination tests, not more demos.