ChinaiFlytek says it trained Spark X2.6 without Nvidia.
Chairman Liu Qingfeng says domestic compute trails Nvidia by nearly 2 generations, in memory and bandwidth.
A 4-billion-parameter on-device model handles easy or sensitive tasks locally, while a cloud model takes the hard ones.
Liu says only Huawei and iFlytek are mainly training on domestic chips, and he wants more firms to join.
iFlytek’s target is 60% on-device task completion, up from 41.9% now.
How each outlet framed it
- TMTPost leans favorable
- details Xunfei's edge-cloud split with a 4B on-device model, 41.9% agent completion vs 12% prior, and the 'two generations behind Nvidia' admission
Sources: TMTPost