iFlytek released Spark X2.6 on October 22 at its 1024 developer festival. It is a 293B-A30B mixture-of-experts model with a 1M-token context window, and iFlytek is pitching it at code and agent work. It is available to try on the Spark website.

Trained on domestic compute, according to iFlytek

iFlytek's Spark API page says Spark-X2.6 was trained on domestic Chinese computing power. At the event, iFlytek chairman Liu Qingfeng went further, saying Spark is the only model trained entirely on domestic computing power while others train on Nvidia GPUs, according to IT Home. That is iFlytek's claim, and we have not verified it.

Limited-time API pricing

The Spark API page lists Spark-X2.6 at a limited-time 40% discount. Per million tokens:

  • Input: 0.9 yuan
  • Cache hit: 0.08 yuan
  • Output: 3.6 yuan

The page gives no end date for these prices. For comparison with other Chinese model APIs, see our DeepSeek API pricing and MiniMax M3.1 Flash pages.

iFlytek Says It Open-Sourced a 4B Model

iFlytek says it has also open-sourced an on-device model, Spark X2.6-4B, and that its inference is free for a limited time. iFlytek did not give an end date for that either. For another recent open on-device release, see MiniCPM-V 4.7.

We have not tested Spark X2.6 or the 4B model.

Sources

  • iFlytek Spark API page
  • IT Home report