On the afternoon of October 22, iFlytek officially released the Spark X2.6 edge-cloud integrated large model at its Global 1024 Developer Festival, claiming the first breakthrough in code and agent capabilities trained on fully domestic computing power.
Point one: parameters and architecture. Spark X2.6 was trained on domestic compute, uses a 293B-A30B MoE architecture, expands context length to 1M, reaches top-tier code and agent capabilities, debuts an edge-cloud collaborative architecture, is now open for trial on the Spark website, and supports custom security policies per task scenario.
Point two: domestic compute breakthrough. Chairman Liu Qingfeng said Spark is the only large model trained on fully domestic compute, and the smallest in size. X2.6 first broke the training-efficiency bottleneck for long-horizon tasks on domestic chips, with dynamic parallel training technology for domestic platforms that solves low mixed-training efficiency caused by memory-bandwidth limits.
Point three: a 4B on-device model open-sourced at the same time. The Spark X2.6-4B open-source on-device model delivers a step change in on-device agent capability, with major gains in code, agents, long text and multilingual ability, three new inference modes, inference efficiency under its own framework better than ollama, and limited-time free inference pricing. Since its predecessor X2.5-4B was open-sourced on September 1, its downloads have ranked among the global top four on-device models.
Source: https://finance.sina.cn/stock/jdts/2026-10-22/detail-iniwarrp5941216.d.html
Comments