inclusionAI releases Ming-3.0-audio-flash: 126B MoE / 6.5B active, audio understanding + speech output, MIT
retained text
[Ming-3.0-audio-flash is up on Hugging Face](https://huggingface.co/inclusionAI/Ming-3.0-audio-flash). 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.
document fetched 2026-10-20 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1x9yrfi/inclusionai_releases_ming30audioflash_126b_moe