MiMo-V2.5-ASR

MiMo-V2.5-ASR is a state-of-the-art end-to-end automatic speech recognition model developed by the Xiaomi MiMo team. Audio first passes through the MiMo-Audio-Tokenizer, a 1.2B-parameter transformer operating at 25 Hz that uses an eight-layer residual vector quantization (RVQ) stack to produce 200 discrete tokens per second; a patch encoder then aggregates four consecutive time steps into a single patch, downsampling the sequence to a compact 6.25 Hz representation. The tokenizer was trained from scratch on a 10-million-hour corpus, jointly optimizing semantic and reconstruction objectives (huggingface.co/XiaomiMiMo/MiMo-Audio-Tokenizer). Those tokens feed an 8B-parameter transformer decoder with Flash Attention support, trained through large-scale mid-training, high-quality supervised fine-tuning, and a novel reinforcement-learning algorithm (huggingface.co/XiaomiMiMo/MiMo-V2.5-ASR). On Xiaomi's published benchmarks the model achieves 2.52% CER on AiShell-2 and 2.41% on Fleurs-Zh for Mandarin -- leading both FunASR-1.5 and Seed-ASR 2.0 -- alongside 5.73% average WER across eight English Open ASR Leaderboard test sets with 1.45% on LibriSpeech-Clean. For Cantonese, the model scores 3.28% CER on Fleurs-Yue and 7.21% CER on a second Cantonese test set (Xiaomi's published MiMo-V2.5-ASR benchmark results). It handles code-switched Mandarin-English speech without language tags, generates punctuation natively from prosody and semantics, and is designed for noisy, multi-speaker, and overlapping-conversation scenarios. This OpenASR pack ships the q4_k build under an MIT license inherited from the upstream model, running entirely on-device via the .oasr format -- no cloud upload, no Python at inference time. OpenASR's catalog entry supports English, Cantonese, and Mandarin Chinese; the base model's broader dialect training on Wu, Hokkien, and Sichuanese provides additional architectural depth but those dialects are not shipped as supported languages in this pack.