MSI-Bench - English: leaderboard

Metric: All-pass rate (%; all 576 English test cases over six multi-speaker memory, instruction-following and reasoning patterns; a case passes when all its atomic rubrics pass, judged by DeepSeek V4 Pro on the dialogue transcript, with a deterministic tool-call validator as one more rubric for cases that need a tool call; one prediction per model and case from the audio of a multi-party, multi-turn scene ending in an assistant-directed handoff; open-weight models served with JSON-schema constrained decoding). Source: arxiv.org. Saturation forecast: Around December 2026. 15 models tracked.

Top models

#ModelScore
1Gemini 3.1 Pro (Preview)66.8
2Gemini 3.5 Flash63.2
3GPT Audio43.9
4Gemma 4 12B (Non-reasoning)22.6
5GPT Audio Mini20.7
6Gemma 4 12B (Reasoning)19.8
7Qwen2.5-Omni-7B16.1
8Phi-4 Multimodal Instruct5.4

Interactive version: theaggregate.ai/benchmark?slug=msi-bench-english · How It Works · Data refreshed daily, snapshot 2026-09-26.