inclusionAI releases five open-weight Ming-3.0 speech and audio models
Article timeline
The articles in this story. Times are ET.
vLLM@vllm_projectOfficialAnt Ling open-sources five Ming-3.0 speech models, with vLLM-Omni day-0 supportAIAnt Ling open-sourced five Ming-3.0 speech models spanning TTS, audio understanding, and ASR, from 0.5B to 126B total parameters (6.5B active). vLLM-Omni supports serving them from day 0.
Ant Ling@AntLingAGIOfficialAnt Ling open-sources five Ming-3.0 speech models across three familiesAIAnt Ling says it has open-sourced five Ming-3.0 models across speech synthesis, audio understanding, and speech recognition. The Ming-3.0-tts family includes a 0.5B voice design model and a 7.9B-parameter (1.3B active) Tiny model for long-form speech. The Ming-3.0-audio family includes a 126B-parameter (6.5B active) Flash model for understanding and tool use, and a 9.3B-parameter (2.7B active) Tiny model for efficient dialogue, while Ming-3.0-asr adds an 8.6B-parameter (2B active) Tiny model for dialects and context.
SGLang@sgl_projectOfficialSGLang-Omni adds day-0 support for Ming-3.0 audio modelsAISGLang-Omni supports Ming-3.0 from InclusionAI on day 0, covering three audio models. Ming-3.0-asr-tiny handles Chinese and English speech recognition with dialect support, and Ming-3.0-tts-tiny clones voices in 7 languages. Ming-3.0-audio-flash is a 126B-A6.5B MoE model for audio understanding and reasoning.
ModelScope@ModelScope2022OfficialMing-3.0 releases five open-weight audio models for speech and transcriptionAIinclusionAI releases Ming-3.0, five open-weight audio models for controllable speech, long-recording analysis, and context-aware transcription. Ming-3.0-tts-0.5B reaches a 1069 Elo rating on Artificial Analysis Controlled Voice, the top open-weight result, using eight English reference voices. Ming-3.0-audio-flash scores 99.40 on BigBenchAudio S2T in the authors' comparison, and Ming-3.0-asr-tiny accepts hotwords and dialogue history.
