inclusionAI releases Ming-3.0-audio-flash: 126B MoE / 6.5B active, audio understanding + speech output, MIT

Ming-3.0-audio-flash is up on Hugging Face . It takes audio alongside text instructions and can return text or a spoken response. The model card lists transcription, spoken Q&A, long-recording understanding, and reasoning about speech, environmental sounds and music. A few details: 126B total / 6.5B active , according to the model card. Chinese and English are the primary languages, with support claimed for accents, dialects and noisy recordings. Optional Talker for speech output, using built-in voices or reference audio. MIT license. For local deployment, the card documents Transformers, SGLang and vLLM-Omni paths. Its validated SGLang setup uses four GPUs for text output; the vLLM-Omni 128K reference profile uses eight 80GB H800s. Those are reference configurations, not minimum requirements. The first thing I'd want to test is a messy recording with people talking over each other, then ask questions about it. That's a more useful test for the kind of voice assistant I'd actually want to run.