iFlytek releases Spark X1.5

Background iFlytek (科大讯飞) develops the Spark (讯飞星火) line of large language models. The release was held at the iFlytek 1024 Developer Festival, the company's developer event. ^[1]

What happened iFlytek released Spark X1.5 on 6 November 2025 at its 1024 Developer Festival. ^[1]

The model uses a mixture-of-experts (MoE) design with 293 billion total parameters. In a design of this kind, only part of the model is active for each input, so the total size and the amount of work per answer are different things. ^[1]

Compared with X1, the model's inference efficiency doubled. Inference efficiency is the amount of useful output a model produces for a given amount of computing. ^[1]

The model supports more than 130 languages. ^[1][2]

Key details iFlytek claimed the model reached 95% of GPT-5's overall level. The figure is the company's own assessment and is reported in the same release as the efficiency and language claims. ^[2]

The company also reported that its open platform had 9.68 million developers by the end of October 2025. The open platform is where developers access the company's models and build applications on them. ^[1]

The claims in this release are the company's own statements. They are useful for understanding what iFlytek was emphasising, namely efficiency, language coverage and a comparison with GPT-5. ^[1][2]

Parameters are the numerical values inside a model that are adjusted during training. The total parameter count is a rough measure of a model's size. In a mixture-of-experts design only part of the parameters is used for each input, so size and the cost of each answer are not the same thing. Inference efficiency describes the work a model does for a given amount of computing, and it is a different measure from parameter count. A reader comparing release claims should keep the two apart.

Impact The release gives a snapshot of the Spark reasoning model at the end of 2025. The points iFlytek emphasised were inference efficiency, a mixture-of-experts design and language coverage. The number of developers on the open platform gives a sense of the community that uses the Spark models. ^[1]