On 2026-10-12 OpenBMB released MiniCPM-V 4.7, an Apache-2.0 open multimodal family with a 35B-A3B variant (the series' first MoE) and a 1B variant. Its Lookahead Expert Scheduling (LES) prefetches experts across layers and, combined with layer-wise mixed-precision quantization and DSpark speculative sampling, lifts SSD-offloaded decoding on an iPad Pro M4 (16GB) from 2.97 to 17.39 tokens/s (5.86x), per OpenBMB. Weights are on Hugging Face, ModelScope and Ollama, with vLLM, SGLang, llama.cpp and iOS/Android/HarmonyOS support.
Key Takeaways
- ✓Two Apache-2.0 open-weight models: 35B-A3B (first MoE in the series, Qwen3.6 language backbone) and 1B, plus FP8 / AWQ / GPTQ / int4 / GGUF builds and DSpark draft models (Hugging Face)
- ✓LES on-device decode (OpenBMB, iPad Pro M4 16GB): baseline 2.97 → LES 9.81 → +layer-wise quant 14.92 → +DSpark 17.39 tokens/s, 5.86x overall (LES doc)
- ✓35B-A3B reaches 13+ tokens/s on iPhone 17 Pro Max; mixed 4x/16x visual-token compression trades detail for cost per task
- ✓Vendor-reported (Instruct, 4x): MMBench 91.72, OCRBench 91.20, OmniDocBench v1.6 94.24, MMMU 80.22 vs Qwen3.6-35B-A3B at 86.76 / 86.00 / 80.17 / 77.44; still behind on SimpleVQA and ScreenSpot
- ✓Day-0 chip support for Rockchip, Houmo, Intel, AMD and Arm; runs on vLLM, SGLang, llama.cpp and Ollama; fine-tune via SWIFT / LLaMA-Factory
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background
On-device multimodal models face a capacity-vs-memory trade-off: MoE models activate few parameters per token but their full expert set rarely fits in phone memory, and naive SSD offloading stalls decoding. OpenBMB's MiniCPM-V 4.7 ships Apache-2.0 weights for a 35B-A3B MoE (first in the series, Qwen3.6 backbone) and a 1B model.
How it works
Lookahead Expert Scheduling predicts next-layer routing to prefetch experts, computes already-loaded experts first and loads the rest in parallel; real routing is unchanged and no retraining is needed. Layer-wise mixed-precision quantization, contiguous expert repacking and DSpark speculative sampling stack on top. Vision uses LLaVA-UHD v4 intra-ViT early compression with selectable 4x/16x visual-token compression.
Numbers (vendor-reported)
iPad Pro M4 16GB decode: 2.97 → 9.81 (LES) → 14.92 (+quant) → 17.39 tokens/s (+DSpark), 5.86x; 13+ tokens/s on iPhone 17 Pro Max. Instruct 4x: MMBench 91.72, OCRBench 91.20, OmniDocBench 94.24, MMMU 80.22 vs Qwen3.6-35B-A3B 86.76 / 86.00 / 80.17 / 77.44; it does not lead on SimpleVQA, ScreenSpot or MM-IFEval.
Getting started
Serve with transformers serve openbmb/MiniCPM-V-4.7-35B-A3B, vLLM or SGLang; run locally via Ollama or GGUF on llama.cpp. Pass downsample_mode to both apply_chat_template and generate. Mobile builds live in MiniCPM-V-Apps; AWQ/GPTQ/int4 variants are on Hugging Face.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.