On October 19 Ant Ling (inclusionAI) open-sourced five Ming-3.0 models under MIT, spanning speech synthesis (tts-0.5B; tts-tiny, 7.9B total/1.3B active), audio understanding (audio-flash, 126B/6.5B active; audio-tiny, 9.3B/2.7B active) and speech recognition (asr-tiny, 8.6B/2B active). Per the vendor, tts-0.5B ranks #1 among open-weight models on Artificial Analysis Controlled Voice at 1069 Elo and ties #1 overall on pronunciation robustness at 93.2%; asr-tiny averages 4.41% error across four ASR categories. Weights are on Hugging Face and ModelScope, with day-0 SGLang and vLLM support.
Key Takeaways
- ✓Lineup: five models in three families, two TTS, two audio-understanding and one ASR, all MIT-licensed.
- ✓TTS results: tts-0.5B scores 1069 Elo on AA Controlled Voice, #1 among open weights; 1188 Elo and #3 on Provider Voice; ties #1 overall on pronunciation robustness at 93.2% (vendor figures).
- ✓TTS features: 7 languages plus Sichuanese and Cantonese, 14 inline paralinguistic tags (laughter, sighs, coughs and more), 44.1 kHz output, built on Qwen2.5-0.5B.
- ✓Audio understanding: audio-flash has 126B total and 6.5B active parameters, and handles long-recording summaries, spoken Q&A, audio reasoning and tool-call generation.
- ✓ASR: asr-tiny accepts hotwords and dialogue context, is evaluated on 15 dialect test sets, and averages 4.41% error across four categories (vendor figures).
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and Pain Point
Voice agents, podcast generation and call-center QA need models that speak, listen and understand, but open options are fragmented, with weak voice control, poor dialect recognition and unclear commercial licenses. Ant inclusionAI released TTS, audio understanding and ASR together under one MIT license, lowering the bar for a self-hosted speech stack.
Architecture Highlights
tts-0.5B is a dense model on Qwen2.5-0.5B with a DiT head, audio tokenizer and speaker encoder, supporting zero-shot cloning, voice design from a text description, emotion/speed/pitch control and 14 paralinguistic tags. tts-tiny, on Ling V3 (7.9B total, 1.3B active), targets long multi-speaker podcasts. audio-flash pairs an audio encoder with an MoE language backbone (126B total, 6.5B active) and an optional Talker for speech replies. asr-tiny, on Ling-tiny-v3, takes hotwords and prior dialogue in its prompt.
Benchmarks
Vendor figures: tts-0.5B scores 1069 Elo on Artificial Analysis Controlled Voice, #1 among open weights; 1188 Elo and #3 on Provider Voice; ties #1 among all models on pronunciation robustness at 93.2%. asr-tiny averages 4.41% error across four ASR categories and is evaluated on 15 dialect sets. A consolidated audio-flash benchmark summary was not given in the launch posts.
Developer Hands-On Guide
Weights live under inclusionAI on Hugging Face and ModelScope. TTS inference requires the Ming-3.0-tts GitHub repo; it was tested on H800 and H20 with CUDA 12.4, and flash-attn is optional. audio-flash and asr-tiny use custom Transformers code with trust_remote_code, so review the code first. For serving, SGLang-Omni and vLLM have day-0 support.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.