Kimi-ASR Internal Testset - Subset 2: leaderboard

Five audio language models transcribe subset 2 of Moonshot's unreleased in-house ASR test set (language and size not stated); the number is word error rate, lower is better. Unlike LibriSpeech or AISHELL nobody outside Moonshot can run it: Moonshot ran all five models itself with Kimi-Audio-Evalkit for the Kimi-Audio report (2025).

Metric: WER (%), lower is better, Moonshot's internal ASR test set subset 2, Kimi-Audio-Evalkit. Source: raw.githubusercontent.com. Saturation forecast: Rough model projection: around 2026. 5 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1Kimi-Audio-7B-Instruct2.44#1517
2Qwen2.5-Omni-7B2.68#1883
3Qwen2-Audio-7B3.24
4Step-Audio-Chat4.74
5Baichuan-Audio-Base5.6

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=kimi-asr-internal-testset-subset-2 · How It Works · Data refreshed daily, snapshot 2026-11-17.