AI Research Atlas

MiniCPM-o 4.5

ModelBest (OpenBMB) · 4 February 2026

9B omni-modal model with full-duplex live streaming of vision and speech; scores 77.6 on OpenCompass vision, above GPT-4o in its own comparison.

Built from SigLip2, Whisper-medium, CosyVoice2 and Qwen3-8B. Apache-2.0. A realtime web demo was open-sourced on 2026-02-06.

Date
Wednesday, 4 February 2026
Lab
ModelBest (OpenBMB)
Kind
open-weights
Access
open weights

Figures

MeasureValueMeasured by
OpenCompass average (8 vision benchmarks)77.6
vs GPT-4o and Gemini 2.0 Pro per the card
company

Date from the HF commit history (2026-02-04 UTC; HF repo created 2026-02-03). Benchmark comparisons are the authors'.

Sources

  1. huggingface.co/openbmb/MiniCPM-o-4_5

This record was checked against its sources on 6 October 2026. How we check