iFlytek launches Spark X2.6 end-cloud model and open 4B edge model
iFlytek released the Spark X2.6 end-cloud model with a 293B-A30B MoE architecture, 1M-token context, and an emphasis on code and agent tasks. It also launched the open-weight Spark X2.6-4B edge model and made the X2.6 series available for trial, with promotional inference pricing of RMB 0.9 per million input tokens and RMB 3.6 per million output tokens. iFlytek says X2.6 was trained entirely on domestic computing infrastructure and that its dynamic parallel training and distributed inference methods address memory-bandwidth constraints, but the reports provide no independent benchmarks for these claims.
The release adds a large end-cloud model and a smaller open edge model focused on coding and agents, while the reported domestic-compute training approach is a significant part of iFlytek's product positioning.
IT之家 10 月 22 日消息,在今日下午的科大讯飞全球 1024 开发者节活动中,讯飞星火 X2.6 端云一体大模型正式发布。


讯飞星火 X2.6 宣称首次基于全国产算力实现代码和智能体能力突破。科大讯飞董事长刘庆峰表示,只有星火一家是用全国产算力训练的,其他全是在 N 卡上训练,而且科大讯飞的尺寸是最小的。
