Ming-3.0

Ming-3.0 is a family of audio and speech models released by InclusionAI. Its official collection includes two audio-language models, a speech recognizer and two speech synthesizers. The audio-language models accept recordings and text instructions, while the specialized models perform speech recognition or text-to-speech.^{[1][2][3][4][5][6]}

Models

The five checkpoints have different architectures, tasks and inference requirements. The parameter figures below retain the scope stated in the publisher's documentation.^{[2][3][4][5][6]}

Model names and documentation figures should be distinguished from counts of tensors distributed in a repository. On November 4, 2026, Hugging Face's Safetensors metadata reported 128,181,010,400 parameters for audio-flash, 8,871,306,016 for audio-tiny, 8,865,800,992 for asr-tiny, 8,981,004,468 for tts-tiny and 1,479,454,868 for tts-0.5B. These repository counts differ from several model-card figures. The TTS checkpoints include components beyond their named language backbones.^{[2][3][4][5][6][7]}

Audio understanding and spoken responses

The audio-flash and audio-tiny models combine an audio encoder and projector with a language backbone. An optional component called the Talker generates spoken responses. Their documented tasks include transcription, spoken question answering, understanding non-speech sounds and music, and reasoning about recordings. Their primary languages are Chinese and English, including accents and dialects.^[2][3]

Speech output can use a built-in voice or a supplied reference voice. This is a separate output stage: the language model can also return text alone. The shared inference repository supplies audio-chat, transcription and speech-output examples.^[8]

The publisher provides serving recipes for vLLM-Omni, SGLang and SGLang-Omni. Their supported outputs and device layouts differ. For example, the documented Flash SGLang profile returns text, while the vLLM-Omni recipes include text-only and Talker profiles. The reference Flash profile for a 128K context uses four 80 GB H800 GPUs; the documentation warns that particularly long audio can require an eight-GPU fallback.^[2][8]

Context-aware speech recognition

Ming-3.0-asr-tiny specializes in single-turn Chinese and English transcription, including regional dialects. Its contextual recognition mode accepts relevant hotwords, a topic description or preceding dialogue in the transcription prompt. These hints can help with names and domain terminology, but do not guarantee that a supplied term will appear.^[4]

The reference implementation downmixes audio and resamples it to 16 kHz. It produces plain transcription text without word-level timestamps or speaker diarization. The included Transformers examples use one CUDA GPU and non-streaming inference. They do not accept images or video.^[9]

The released configuration sets an 8,192-token context. Longer recordings can be encoded in chunks within the model's context limit. The Context ASR example provides separate handling for known-language clips up to 60 seconds. SGLang-Omni can stream transcript output as server-sent events; this differs from the reference example's lack of streaming audio input.^[4][10]

Speech synthesis

Both TTS models support Chinese, English, German, Spanish, French, Korean and Japanese, with Sichuanese and Cantonese dialect controls. They produce audio at 44.1 kHz. Their shared generation framework conditions a language model on text and control tokens, then uses a diffusion-transformer patch head and an audio tokenizer to generate speech.^[5][6][11]

The available controls include reference-voice cloning, designing a voice from a text description, choosing built-in voices, selecting emotion and adjusting speed, pitch and volume. Fourteen inline paralinguistic tags specify events such as laughter, coughing and sighing in Chinese or English. These features use task-specific modes with different accepted inputs.^[5][6]

Reference requirements depend on the model and task. For tts-tiny, ordinary reference-conditioned synthesis, multilingual synthesis and paralinguistic synthesis require reference audio and its transcript. Its podcast task requires a recording and matching transcript for each speaker, with every script line identifying its speaker. Voice design and built-in voices require no reference recording.^[5]

The two TTS models also differ in free-form control. Tiny requires a reference clip for its freeform task, while the 0.5B model can follow a reference voice or choose a voice through its instruction when no reference is supplied. Free-form control cannot be combined arbitrarily with the other control modes.^[5][6]

Inference environments

The Hugging Face TTS repositories distribute checkpoints; inference uses the separate Ming-3.0-tts code repository. The 0.5B instructions document tests on NVIDIA H800-80GB and H20-96G with CUDA 12.4. FlashAttention is optional, with a PyTorch scaled-dot-product-attention fallback.^[6]

Tiny has a separate reference environment: Python 3.10.15, PyTorch and torchaudio 2.7.1 with CUDA 11.8 wheels, and Transformers 4.46.3. Its vLLM-Omni recipe uses another environment. SGLang-Omni also supplies a speech endpoint. Dependency and serving instructions should therefore be taken from the selected checkpoint's documentation.^[5]

Evaluation and limitations

The audio-language project publishes speech-recognition, audio-understanding and text benchmark figures. It states that Flash was evaluated without thinking mode and that comparison results can differ from official reports because of prompts, decoding settings, versions and evaluation pipelines.^[8]

The TTS project's published Seed-TTS evaluation gives the following results. Word error rate (WER) measures transcription errors in synthesized speech, with lower values preferred; speaker similarity (SIM) measures similarity to the reference voice, with higher values preferred. These are the project's reported results for the specified evaluation.^[11]

The TTS documentation identifies weaker speaker similarity in several multilingual evaluations, difficulty controlling multiple vocal events in one sentence, and less accurate pitch control than speed or volume control. Tiny's documented limitations include Korean content accuracy and podcast audio quality. The project also reports difficulty expressing an emotion when the words provide no corresponding emotional signal.^[5][6]

License and availability

The five model cards identify the MIT license. The ASR project specifies that its original code and weights use MIT, while bundled Apache-2.0 files retain their own notices. Its demo recordings retain their source licenses.^{[2][3][4][5][6][9]}

The TTS project's usage guidance calls for consent when cloning a real person's voice and disclosure where synthetic speech could be mistaken for a real recording. These statements accompany the project's voice-cloning and built-in-voice features.^[11]

Continue reading

• SSpeech RecognitionSpeech recognition, usually called automatic speech recognition (ASR), is the computational task of converting a spoken-language signal into a sequence of written symbols.Mentioned twice above
• InclusionAIinclusionAI is an open-source artificial general intelligence (AGI) research initiative established by Ant Group, the financial-technology affiliate of the Alibaba ecosystem.Linked in the introduction
• TText-to-SpeechText-to-speech (TTS) refers to artificial intelligence systems that convert written text into natural-sounding spoken audio.Linked in the introduction
• Qwen2-AudioQwen2-Audio is an audio-language model developed by the Qwen team at Alibaba Cloud, released in August 2024.Also in Speech & Audio AI
• CartesiaCartesia is a San Francisco-based AI company focused on real-time voice synthesis, speech recognition, and state space model (SSM) research.Also in Speech & Audio AI
• AAudio ModelsAudio models are machine learning systems that take audio as input, produce audio as output, or both, spanning speech recognition, speech synthesis, music generation, sound effect generation, voice conversion…Also in Speech & Audio AI

More to read

• Lyria
• Automatic Speech Recognition Models
• Text-to-Speech Models
• ElevenLabs v3
• Moshi
• Nemotron 3.5 Transcribe

Topics:AI ModelsMultimodal AISpeech & Audio AI

References

1. ^inclusionAI. Ming-3.0 collection. Hugging Face, accessed November 4, 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7inclusionAI. Ming-3.0-audio-flash model card. Hugging Face, accessed November 4, 2026.
3. ^1 ^2 ^3 ^4 ^5 ^6inclusionAI. Ming-3.0-audio-tiny model card. Hugging Face, accessed November 4, 2026.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7inclusionAI. Ming-3.0-asr-tiny model card. Hugging Face, accessed November 4, 2026.
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11inclusionAI. Ming-3.0-tts-tiny model card. Hugging Face, accessed November 4, 2026.
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10inclusionAI. Ming-3.0-tts-0.5B model card. Hugging Face, accessed November 4, 2026.
7. ^Hugging Face. Safetensors metadata for audio-flash, audio-tiny, asr-tiny, tts-tiny and tts-0.5B. Model API, retrieved November 4, 2026.
8. ^1 ^2 ^3inclusionAI. Ming-3.0-audio-flash: architecture, benchmarks and inference. GitHub, accessed November 4, 2026.
9. ^1 ^2inclusionAI. Ming-3.0-asr: inference and usage notes. GitHub, accessed November 4, 2026.
10. ^inclusionAI. Ming-3.0-asr-tiny configuration. Hugging Face, accessed November 4, 2026.
11. ^1 ^2 ^3inclusionAI. Ming-3.0-tts: architecture, evaluation and usage guidance. GitHub, accessed November 4, 2026.

Show all 11 references