Long recordings are where most speech pipelines fall apart: you transcribe, chunk, summarise and hope nothing important was lost between stages. This model takes the other route. It listens to the audio directly and answers questions about it, keeping cues such as emotion, speaker identity and background sound that a plain transcript throws away.
It is the largest member of inclusionAI's Ming-3.0 speech series and ships with downloadable weights under the MIT licence.
Under the hood
The design pairs an audio encoder and projector with a sparse mixture-of-experts language backbone drawn from Ling-3.0, plus an optional "Talker" module that turns the response into speech. Input audio is framed at 6.25 Hz, which keeps long clips affordable in tokens. The attention stack is a hybrid of KDA and gated MLA layers.
| Item |
Detail |
| Parameters |
126B total, 6.5B active (file metadata rounds to 128B) |
| Inputs |
Audio, text |
| Outputs |
Text; speech via the optional Talker |
| Languages |
Primarily Chinese and English |
| Licence |
MIT |
| Hosted API |
Ant Ling open API, from 19 October 2026 |
What it can actually do
- Speech recognition across dialects, accents, noisy rooms and multi-speaker audio
- Spoken question answering and reasoning over what was said
- Long-audio summarising and viewpoint analysis with structured output
- Recognition of non-speech sounds, acoustic scenes and music
- Spoken replies in a built-in voice or a reference voice
Running it yourself
Plan for real hardware. The reference vLLM-Omni profile uses tensor parallelism across four 80 GB H800 GPUs, with an eight-GPU fallback for very long audio, since long inputs can exhaust memory on the four-card setup. SGLang has a validated four-GPU profile for text output only; SGLang-Omni runs on two GPUs and exposes OpenAI-compatible endpoints. Loading in Transformers needs trust_remote_code=True, so review the repository code before you run it.
Dating is slightly messy. The repository was created in early September 2026, while the model card and hosted API arrived on 19 October. If you are pinning a version, pin the commit, not the date.
Picking a test set
Build your trial from recordings you already own: a noisy call, a two-hour meeting, a dialect sample. Ask the same factual questions you would ask a human note-taker and grade the answers, then compare with a transcribe-then-summarise chain. If you only need text out, the smaller Ming-3.0-asr-tiny is far cheaper to host; for voice output alone, see Ming-3.0-tts-0.5B. The broader audio-text-to-text list shows other options, and the inclusionAI page collects the rest of the family.