126B total / 6.5B active (Flash); 9.3B total / 2.7B active (Tiny), both MoE
131,072 tokens (128K text configuration)
MIT
Open-weight native audio-language family comprising Ming-3.0-audio-flash and Ming-3.0-audio-tiny. Both accept audio together with text instructions and produce text or optional synthesized speech, supporting automatic speech recognition, spoken question answering, long-recording analysis, non-speech sound and acoustic-scene understanding, music understanding, audio reasoning, spoken interaction, and audio-conditioned tool calling. Flash uses a 42-layer Mixture-of-Experts language backbone with a 2,560-dimensional hidden state, 32 attention heads, 512 routed experts with eight selected per token, and one shared expert; inclusionAI advertises 126B total parameters with 6.5B active per inference step. Tiny uses 24 language layers, a 1,536-dimensional hidden state, 16 attention heads, 128 routed experts with eight selected, and one shared expert, advertised as 9.3B total and 2.7B active. Both attach a 32-layer, 1,280-dimensional audio encoder using 128-bin mel features and three stride-two downsampling stages. Their released configurations specify 131,072 text positions. Optional Talker and VAE components provide built-in voices and reference-audio voice prompting. Hugging Face safetensor metadata enumerates approximately 128.18B parameters for the Flash main checkpoint and 8.87B for Tiny, alongside separately packaged speech-generation assets. Official inference paths include custom Transformers code, vLLM-Omni, and SGLang. Primary recognition and interaction support is documented for Chinese and English, including accents and Chinese dialects. Publisher-reported Flash/Tiny results include 99.40/99.20 on BigBenchAudio, 82.00/75.90 on MMAU, 68.41/64.54 on AudioMarathon, 86.74/74.29 on VoiceBench, 81.40/79.80 on BFCL-v3 Audio, Chinese average CER of 3.45/3.79, English average WER of 3.69/4.36, and Chinese-dialect average CER of 9.40/10.29. The Flash text backbone additionally reports 82.45 on GPQA Diamond and 79.90 on LiveCodeBench v6. The broader same-day Ming 3.0 speech release includes three specialized sibling checkpoints that are documented here but not counted as part of this card's parameter or context fields. Ming-3.0-tts-0.5B is a dense Qwen2.5-0.5B-based speech synthesizer, while Ming-3.0-tts-tiny uses a 7.9B-total / 1.3B-active Ling V3 MoE backbone. Both generate 44.1-kHz speech across Chinese, English, German, Spanish, French, Korean, and Japanese and support zero-shot voice cloning, natural-language voice design, free-form delivery control, emotion, speed, pitch, volume, Sichuanese and Cantonese, and inline breathing, coughing, sighing, laughter, hesitation, and other paralinguistic cues. Ming-3.0-asr-tiny is an 8.6B-total / 2B-active contextual transcription model for Chinese, English, and regional speech, evaluated across fifteen Chinese dialect sets and accepting hotwords or preceding dialogue to improve domain-term recognition; inclusionAI reports a 4.41% average error across four ASR categories. All five checkpoints are released under MIT.