MiniCPM-o-4.5: benchmark results

An end-to-end multimodal checkpoint combining SigLIP2, Whisper-medium, CosyVoice2 and Qwen3-8B, with a publisher-rounded total of 9B parameters. It supports full-duplex audio/video streaming and both instruct and thinking modes. Provider: OpenBMB. Released 2026-02-03. Access: Open.

Aggregate Rating 1589 ± 15, rank #1048 of 10870 rated models, from 300 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
CUE-Mem - Multimodal Index Multimodal Use - Answer Refusal Explicit50.3Accuracy (%)100
CUE-Mem - Multimodal Index Text Use - Answer Refusal Explicit53.8Accuracy (%)100
CUE-Mem - Multimodal Index Text Use - Long Pattern Implicit39.4Accuracy (%)100
CUE-Mem - Oracle Index Text Use - Entity Recall Explicit91.7Accuracy (%)100
CUE-Mem - Oracle Index Text Use - Long Pattern Explicit81.2Accuracy (%)100
CUE-Mem - Oracle Index Text Use - Memory Average Explicit82.6Accuracy (%)100
CUE-Mem - Text Index Text Use - Answer Refusal Explicit49.2Accuracy (%)100
CUE-Mem - Text Index Text Use - Answer Refusal Implicit89.8Accuracy (%)100
DuplexSpeechBench-IFEval (PAS)66.2Persona Adherence Score (0-100; GPT-4o-judged persona-consis100
HUG-VIS - Emotion Recognition (Video)48.52Seven-class emotion recognition accuracy (%; 8,400 seated ha100
KMMAU - Topic Summary98Accuracy (%; four-choice question on the topic of the audio;100
MOV-Bench - Hypothetical Reasoning40Accuracy (%; 60 hypothetical-reasoning questions; direct inf100

Interactive version: theaggregate.ai/model?slug=minicpm-o-4-5 · How It Works · Data refreshed daily, snapshot 2026-11-12.