iFlytek trained a 293-billion-parameter model without Nvidia chips.
Chinese speech-AI firm iFlytek launched Spark X2.5 on Monday, trained end to end on domestic hardware. The mixture-of-experts model holds 293 billion parameters but runs only 30 billion at a time. It is live on the company's cloud platform, and two small open-source versions arrived last week.
Why it mattersAmerican export controls were meant to stop exactly this kind of model from being built in China. If Chinese chips can now train large models, the price of AI tools you use may keep falling.
Read in the app — free, in 9 languages
Related stories
Outside testers scored OpenAI's new Astra below Anthropic's Fable 5.1.2026-09-05A Chinese lab built Saudi Arabia's flagship Arabic AI model.
2026-09-04CrowdStrike built an attack AI that hunts its own defence AI.
2026-09-04Google's cheapest new model doubles in price on 1 January.
2026-09-04OpenAI shipped the first model it calls critical for cyber risk.
2026-09-04