iFlytek opens Spark X2.5 with 293 billion parameters on its own platform
iFlytek has released Spark X2.5, a mixture-of-experts model with 293 billion parameters that activates 30 billion parameters per token, on its MaaS platform. The company says the entire training and inference pipeline ran on Chinese hardware and that the model supports a 256K context window and more than 200 languages. Three separate publishers report its coding and agent-performance claims, but no independent test verifies them.
Artificial Intelligence··Evening
Thirty billion of 293 billion parameters activate
iFlytek made Spark X2.5 available on its MaaS platform on September 7. Jiemian News, IT Home and The Yangtzeer describe the same release: the mixture-of-experts architecture contains 293 billion parameters and activates 30 billion for each token. The company says the entire chain, from training to online inference, ran on computing hardware made in China. All three publishers present that as iFlytek's statement; none provides an outside technical examination of the hardware pipeline.[1], [2], [3]
A 256K context window and published API prices
IT Home reports that the flagship model supports a 256K context window and more than 200 languages. iFlytek lists prices of 1.6 yuan per 1 million input tokens, 0.24 yuan per 1 million cached input tokens and 6 yuan per 1 million output tokens. This model is separate from the open-source 4B and 1.7B edge versions released on September 1. Jiemian News and The Yangtzeer say those smaller versions support contexts of up to 1 million tokens, with weights available on Hugging Face and GitHub.[2], [1], [3]
The limit of the performance claims
iFlytek says Spark X2.5 is stronger particularly at coding and agent tasks, with further gains in mathematics and general understanding. Jiemian News attributes those claims to the company's announcement, while IT Home and The Yangtzeer also report the coding and agent focus. None of the sources publishes an independent comparison, reproducible evaluation or third-party test. The architecture, access route and prices can therefore be read as concrete product information, while claims about relative performance currently rest on the vendor's measurements.[1], [2], [3]
