Features
Transcribes Mandarin and English
Supports eight Chinese dialects
Handles code-switched speech
Transcribes song lyrics in mixed vocal/instrumental scenarios
Robust in strong noise, far-field, and multi-speaker conditions
Accurate recognition of knowledge-dense content
Open-source 8B parameter model
Priced at ¥0.5 per hour of audio input
Part of MiMo-V2.5 series (V2 deprecated as of June 30, 2026)
API access via OpenAI/Anthropic compatible endpoints
No real-time streaming (latency not specified)
Does not support speaker diarization
Does not support emotion recognition
Model weights available for self-hosting
Supports audio input only (no text/image generation)
Multilingual speech transcription (99+ languages)
To-English speech translation
Zero-shot robustness to accents, noise, technical language
Open-source models and inference code
Encoder-decoder Transformer architecture
Trained on 680,000 hours of diverse data
Log-Mel spectrogram input
30-second audio chunk processing
Multiple model sizes (tiny to large)
Whisper.cpp for CPU inference
Fine-tuning via Hugging Face integration
Turbo model on OpenAI API
OpenAI API at $0.006 per minute
