iFlytek releases Spark X2.6, trained on domestic chips with a million-token window

On October 22, iFlytek launched Spark X2.6 at its 1024 Global Developer Festival in Beijing, together with the on-device Spark X2.6-4B model.

The cloud model is a mixture-of-experts system: 293 billion parameters in total, about 30 billion active per token, and a context window of 1 million tokens. iFlytek says this is the first time it has reached this level of coding and agent ability on fully domestic computing power. Chairman Liu Qingfeng said long-context training efficiency rose about fourfold and inference efficiency about 4.5 times. Those figures are the company's own, not an independent benchmark.

The 4B on-device model is open source and free for a limited time. The promotional cloud API price is 0.9 yuan per million input tokens, 0.08 yuan on a cache hit, and 3.6 yuan per million output tokens. Liu estimated that about 60 percent of future applications could stay on device, 30 percent would split work between device and cloud, and 10 percent would need the large cloud model.

Next to this week's model news from US labs, Spark X2.6 is mainly about where the model runs: domestic chips, with sensitive work kept on the device when possible.