SSyncDev

Ming-3.0-asr-tiny

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

Ming-3.0-asr-tiny is inclusionAI's open speech-recognition model for Chinese and English with hotword hints, dialect coverage and context-aware ASR.

Key specifications

Specification score
9.8
License
Open weights

The specification score summarizes recorded features and limits. It is not a benchmark result.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Audio input

Produces

  • Text output

Capabilities

AudioOpen weights

What is Ming-3.0-asr-tiny?

Generic speech recognisers stumble on exactly the words a business cares about: product names, people, street names, jargon. This model's answer is to let you hand it those words at inference time. Pass a topic and a hotword list, optionally with preceding dialogue turns, and it uses them as hints while transcribing. No fine-tuning is required.

It is the recognition member of inclusionAI's Ming-3.0 speech series, MIT-licensed, and served through Ant Ling's open API from 19 October 2026.

Size, and a naming quirk

"Tiny" is relative. File metadata puts the checkpoint at roughly 9B parameters, and one description of the release labels it 8.6B with about 2B active. The model card itself does not state an active count, so if memory planning matters, measure the loaded footprint rather than trusting either label.

Where it fits

Item Detail
Backbone Ling-tiny-v3 language model plus an audio encoder adapted from Whisper
Languages Chinese and English
Dialects Evaluated on 15 regional sets, including Cantonese, Minnan and Shanghainese
Output Plain transcript; no word timestamps or speaker diarisation
Licence MIT (bundled Apache-2.0 files keep their own licence)

The missing timestamps and diarisation matter. If you need subtitles or "who said what", you will have to pair it with separate tools, or look at the larger Ming-3.0-audio-flash, which handles multi-speaker audio inside one model.

Deployment notes

Setup targets Linux with Python 3.12 and a CUDA GPU; the card lists Transformers 5.12.1, fla-core 0.5.2 and PyTorch 2.11.0. You can run the bundled examples/transcribe.py through Transformers with trust_remote_code=True, or serve it on a single GPU with SGLang-Omni, which exposes an OpenAI-compatible /v1/audio/transcriptions endpoint with optional streaming. Long audio can be encoded in chunks, within the model's context limit.

Treat hotwords as hints. The card is explicit that supplying a term does not guarantee it appears in the output.

A practical trial

  1. Collect an hour of your own audio, split between clean speech, noisy speech and any dialects your users speak.
  2. Transcribe it with and without a hotword list built from your product catalogue.
  3. Measure word error rate overall and, separately, the hit rate on the hotwords themselves.

That second number is usually the one that justifies a switch. Compare results with other entries on the speech-to-text page before deciding, and see the inclusionAI provider page for the rest of the Ming-3.0 series.

How to evaluate Ming-3.0-asr-tiny

Compare the model on the work you need to complete.

Tasks and inputs

Ming-3.0-asr-tiny is listed for Audio, Automatic speech recognition, Speech → Text. It accepts audio and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under the publisher’s model licence. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Chinese and English, with evaluation across 15 regional dialect sets including Cantonese, Minnan and Shanghainese.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

ElevenLabs

Scribe v3

Scribe v3 is ElevenLabs' batch speech-to-text model, cutting word errors about 27% versus Scribe v2 across 102 languages, at $0.20 per audio hour.

— context

ElevenLabs

Scribe v2 Medical

Scribe v2 Medical is ElevenLabs' HIPAA-eligible batch transcription model for clinical audio, with 35% fewer errors than Scribe v2 on medical speech.

— context