OpenBMB published MiniCPM-V 4.7 weights, model cards and deployment code.
OpenBMB releases MiniCPM-V 4.7 for on-device multimodal inference
2026-10-07Verified announcementTRIAL
OpenBMB released Apache-2.0 weights for MiniCPM-V 4.7 in 35B-A3B and 1B variants. The 35B mixture-of-experts model combines lookahead expert scheduling, quantization and speculative sampling for on-device multimodal inference. The model card reports more than 13 tokens per second on an iPhone 17 Pro Max with 3 GB of memory, but the performance and quality comparisons are author-reported.
What this means for you
For mobile AI teams, public weights and deployment paths make a relatively large multimodal MoE directly testable on phones and local devices. Expert streaming can hide memory pressure without eliminating storage or bandwidth cost, so validate visual quality, sustained speed, thermals and battery use on target hardware.
Evidence & related material
Technical details & scores
| Variants | 35B total / 3B active MoE and 1B dense |
|---|---|
| License | Apache-2.0 |
| Techniques | lookahead expert scheduling, layer-wise quantization and DSpark speculative sampling |
| Platforms | iOS, Android, HarmonyOS, SGLang, vLLM, llama.cpp and Ollama paths |
| Vendor Result | over 13 tokens/s with 3 GB memory on iPhone 17 Pro Max |
| Limits | quality and device-performance comparisons are author-reported |
| Time Basis | 35B repository created 2026-10-07T07:07:30Z and materially updated 2026-10-13T12:47:10Z |
Technical impact 4
Engineering value 4
Adoption signal 2
Maturity 3
Verification cost 3
Risk 2