inclusionAI Releases Ming-3.0-audio-tiny Native Audio-Language MoE Model
inclusionAI has released Ming-3.0-audio-tiny, a compact native audio-language Mixture-of-Experts model featuring 9.3 billion total parameters with 2.7 billion active parameters per inference step. The model supports speech recognition, audio understanding, reasoning, and optional speech synthesis in both Chinese and English. This open-weights release expands the availability of efficient audio-language intelligence within the open-source community, enabling developers to build applications requiring robust speech recognition, multi-speaker handling, and real-time audio interaction. Its Mixture-of-Experts architecture offers high performance while maintaining a relatively low active parameter count. Ming-3.0-audio-tiny accepts both audio and text instructions, yielding text or text with optional synthesized speech using built-in or reference voices. It is licensed under the MIT license and is optimized to handle complex challenges such as noisy environments, varied accents, dialects, and long-form audio recordings.
## BACKGROUND
A native audio-language model is designed from the ground up to process acoustic inputs directly alongside text rather than relying solely on cascaded automatic speech recognition systems. The Mixture-of-Experts (MoE) architecture routes inputs to a subset of neural network subnetworks (experts) per token, allowing models to scale up total capacity while keeping computational costs manageable during inference.