MiMo Audio 7B Base
MiMo Audio 7B Base is a downloadable audio-language model trained at large scale. Xiaomi studies whether broad audio pretraining can produce learning from a few examples without a separate fine-tune for every task.
Research release
The base model is the object of the few-shot study
Xiaomi reports pretraining on more than 100 million hours of audio. It evaluates tasks beyond the standard speech-recognition and synthesis measures, including tasks absent from its training data.MiMo-Audio-7B-Base official model card
MiMo-Audio-Tokenizer is a separate 1.2-billion-parameter component trained from scratch on ten million hours of audio. It operates at 25 Hz and uses eight residual vector-quantization layers to produce 200 tokens per second. Xiaomi publishes its weights alongside the audio-language models; it is the speech representation component, not a separate conversational assistant.MiMo Audio tokenizer architecture and weights
Do not mistake continuation for dialogue alignment
Practical guidance. Use the base checkpoint when studying how audio examples change behavior. For a spoken assistant, compare the instruction-tuned release. Keep examples and target audio separate so the evaluation does not accidentally reward copying.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | MiMo-Audio-7B-Base |
| Research focus | Speech continuation, audio understanding and example-based transfer to tasks such as voice conversion and speech editing.MiMo-Audio-7B-Base official model card |
| Licence | MIT.MiMo-Audio-7B-Base official model card |
| Companion component | The release uses MiMo-Audio-Tokenizer.MiMo-Audio-7B-Base official model card |
Related models
- MiMo Audio 7B Instruct
Xiaomi · MiMo Audio
An instruction-tuned audio model for spoken dialogue, audio understanding and text-to-speech.
Launch date not confirmedDownloadable - MiMo V2.5 ASR
Xiaomi · MiMo V2.5 Speech
V2.5 ASR targets transcription across Mandarin, English, Chinese dialects, mixed-language speech and noisy recordings.
Launch date not confirmedAPI · Downloadable