Physix Frontier · News Briefing Card (IT Home · Sep 16, 2026)

iFlytek Launches Spark-Audio-1.0 on Domestic Compute

KEY FACTS

iFlytek has released the Spark-Audio-1.0-Preview, trained entirely on a domestic computing cluster.
The model features a 0.65B audio encoder and a 30B-A3B MoE architecture, utilizing 13 million hours of audio data.
It supports recognition in 99 languages and 202 dialects, with end-to-end capabilities for understanding semantics, emotions, and environmental sounds.
The model achieved SOTA on the Fleurs Chinese test set, outperforming Gemini-3.1 Pro in multiple speech evaluation metrics.
Compared to competitors of similar size, it excels in multilingual recognition without showing significant performance drops in general knowledge tasks.

KEY DATA

0.65BAudio Encoder Params

30B-A3BLLM Structure Params

13 million hoursTraining Audio Duration

99Supported Languages

PHYSIX OBSERVATION

iFlytek's move signals that China's AI infrastructure is transitioning from merely 'usable' to truly 'effective.' Full-stack domestic training not only mitigates risks associated with compute supply disruptions but also addresses information loss inherent in traditional cascaded solutions through an end-to-end architecture. Its leading performance in real-world scenarios, such as high-noise environments, proves that domestic models possess the practical potential to replace top-tier international closed-source models, offering a cost-effective new option for vertical industry deployment.

Source: IT Home report