openbmb/MiniCPM-V-4.7-35B-A3B
With our Lookahead Expert Scheduling (LES) technique introduced alongside MiniCPM-V 4.7, combined with fine-grained layer-wise quantization and DSpark speculative sampling integration, MiniCPM-V 4.7 35B-A3B becomes the first open-source MoE model to run smoothly on-device without compromising performance, achieving a decoding speed of 13+ tokens/s on iPhone 17 Pro Max.
Proof1 community thread
Key points
- "MiniCPM-V 4.7 is the latest model in the MiniCPM-V series, available in a 35B-A3B variant, with 35B-A3B being the first MoE model in the MiniCPM-V series.
- Notable features of MiniCPM-V 4.7 include:
- The 35B-A3B variant of MiniCPM-V 4.7 advances the performance–efficiency Pareto frontier for MLLMs. In Instruct mode, 35B-A3B achieves a higher overall score in less inference time than Qwen3.6 35B-A3B.
- Based on the intra-ViT early compression technique from LLaVA-UHD v4, MiniCPM-V 4.7 compresses visual tokens early within the vision encoder, reducing computation in subsequent encoder layers.
Sources (1)
- [1]openbmb/MiniCPM-V-4.7-35B-A3Br/LocalLLaMA (top, daily) · Oct 12, 05:16 PM
With our Lookahead Expert Scheduling (LES) technique introduced alongside MiniCPM-V 4.7, combined with fine-grained layer-wise quantization and DSpark speculative sampling integration, MiniCPM-V 4.7 35B-A3B becomes the first open-source MoE model to run smoothly on-device without compromising performance, achieving a decoding speed of 13+ tokens/s on iPhone 17 Pro Max.
"MiniCPM-V 4.7 is the latest model in the MiniCPM-V series, available in a 35B-A3B variant, with 35B-A3B being the first MoE model in the MiniCPM-V series.
Extractive summary: sentences quoted from the sources.
Hype check
Independently evaluated2 independent looks since
- Evaluation · Oct 12, 2026Qwen3.8-27B vs Frontier vs Others
- Evaluation · Oct 12, 2026PhysCodeGym: Benchmarking Coding Agents for Physically Verifiable 4D Reconstruction
- Follow-up · Oct 12, 2026Fix plastic skin: KREA 2 Turbo and I2I workflows for realism
- Follow-up · Oct 12, 2026Openbroca
- Follow-up · Oct 12, 2026Can we change an AI model’s architecture without retraining it from scratch?
- Follow-up · Oct 12, 2026Comprehensive Workflow for Uncertainty-Aware Ionic Liquid Solvent Discovery
Before this
- Oct 12, 2026Z.ai GLM-Legal Beats Claude Fable 5 and Kimi K3 on Harvey LAB
- Oct 12, 2026Deploy decision models with llama.cpp on Amazon SageMaker AI
- Oct 12, 2026unslothai/unsloth v0.1.906-beta: Browser Use + Qwen-Image-2.1-Turbo
- Oct 12, 2026Qwen3.8-Flash-Next on six 3090s with vLLM: 7 users at 262k, 6.5k t/s prefill, 57 t/s decode solo / 206 t/s aggregate with all 7 generating. Fixes for the upstream included.
- Oct 11, 2026Gemma4 E4B EXL3 - Alternative to Llama.cpp on Jetson Orin
- Oct 11, 2026Agents That Own Their Inference — Du'an Lightfoot & Khaja Omer, Akamai Technologies
