HOT RADAR
llmHEAT 11.8°

AI CLUSTERED EVENT · 10/13/2026

OpenBMB Releases MiniCPM-V 4.7: First On-Device MoE Multimodal Model Achieving 13+ Tokens/s on 3GB RAM

1 reports archived1 independent sourcesupdated 10/13/2026, 17:12:15
Synthesis & Latest Updates
1 Sources Cross-Validated

OpenBMB released MiniCPM-V 4.7, a new on-device multimodal LLM available in 35B-A3B and 1B variants. The 35B-A3B version is the first Mixture-of-Experts model in the series, leveraging LES inference acceleration to achieve over 13 tokens/s decoding speed with only 3GB of memory usage, drastically lowering deployment barriers for high-performance models. Additionally, OpenBMB open-sourced the Self-Replay Fine-Tuning (SRFT) method and ensured compatibility across major chip platforms including Intel, AMD, Arm, and Rockchip.

LATEST/OpenBMB released MiniCPM-V 4.7, a new on-device multimodal LLM available in 35B-A3B and 1B variants. The 35B-A3B version is the first Mixture-of-Experts model in the series, leveraging LES inference acceleration to achieve over 13 tokens/s decoding speed with only 3GB of memory usage, drastically lowering deployment barriers for high-performance models. Additionally, OpenBMB open-sourced the Self-Replay Fine-Tuning (SRFT) method and ensured compatibility across major chip platforms including Intel, AMD, Arm, and Rockchip.

HEAT TRENDHourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. IT之家 · 智能时代T2·86 pts
    • MiniCPM-V 4.7 offers 35B-A3B and 1B variants, with the former being the first MoE architecture in the series.
    • Leveraging LES inference acceleration, the 35B-A3B version achieves 13+ tokens/s decoding speed using only 3GB of RAM.
    • The Self-Replay Fine-Tuning (SRFT) method is open-sourced to enable low-cost capability expansion and reduce catastrophic forgetting.
    • Broad compatibility with Intel, AMD, Arm, and Rockchip platforms supports deployment on both iOS and Android devices.