A large model trained entirely on domestically produced chips is here! With a million-token context, its price is even lower than DeepSeek V4.1 Flash.

What happened: iFlytek released the Spark X2.6 cloud-edge model, trained entirely on domestic compute, alongside SparkAudio 1.0, the WinWork platform, AI hardware, and education, medical, and legal assistants. Key numbers: Cloud model 293-A30B supports million-token context at 0.9 yuan per million input tokens, 0.08 yuan cached input, 3.6 Why it matters: iFlytek positions full-stack self-controlled cloud-edge models as the answer to AI security and cost, targeting sectors sensitive to data leaving premises. It says input cache-miss and output prices are lower than DeepSeek-V4.1-Flash-1017. Coverage: 3 outlets in 24h

iFlytek released the Spark X2.6 cloud-edge model, trained entirely on domestic compute, alongside SparkAudio 1.0, the WinWork platform, AI hardware, and education, medical, and legal assistants.

iFlytek positions full-stack self-controlled cloud-edge models as the answer to AI security and cost, targeting sectors sensitive to data leaving premises. It says input cache-miss and output prices are lower than DeepSeek-V4.1-Flash-1017.