inclusionAI Releases Ming-3.0-audio-flash Audio-Language Model
Start here
inclusionAI released Ming-3.0-audio-flash, a native audio-language Mixture-of-Experts model with 126B total parameters and 6.5B activated per inference step, under the MIT License. It takes audio plus text instructions and outputs text or optional synthesized speech, covering speech recognition, spoken question answering, long-audio understanding, audio reasoning and voice interaction in Chinese and English. The model reports 99.8% on Artificial Analysis's Big Bench Audio benchmark, a figure from inclusionAI's own release. The repository documents Transformers, vLLM-Omni and SGLang serving setups. The reference 128K serving profile uses tensor parallelism across four 80 GB H800 GPUs; inclusionAI notes that particularly long audio inputs may run out of memory on the four-GPU profiles, in which case an eight-GPU fallback profile is provided. A two-GPU serving configuration places the thinker on both visible devices, the audio encoder on GPU 0 and speech synthesis on GPU 1.
Generated from reports · updated 21 min ago
Report timeline
Follow the reports to see the event from different sides.
- inclusionAI ModelsPicksinclusionAI releases Ming-3.0-audio-flash, a 126B MoE audio-language model under MIT
inclusionAI released Ming-3.0-audio-flash, a native audio-language Mixture-of-Experts model with 126B total parameters and 6.5B activated per inference step, under the MIT License. It accepts audio plus text instructions and outputs text or optional synthesized speech, covering speech recognition, spoken question answering, long-audio understanding, audio reasoning and voice interaction in Chinese and English. The model reports 99.8% on Artificial Analysis's Big Bench Audio benchmark, and the repository documents Transformers, vLLM-Omni and SGLang serving setups.
Heat for this event
Not enough continuous observations yet to draw a trend.