IT Home, October 22 — At today's afternoon event for the iFlytek Global 1024 Developer Festival, the iFlytek Spark X2.6 end-cloud integrated large model was officially launched.
iFlytek Spark X2.6 claims to be the first to achieve breakthroughs in code and agent capabilities using entirely domestic computing power. Liu Qingfeng, Chairman of iFlytek, stated that Spark is the only model trained on domestic computing power, while all others were trained on NVIDIA GPUs, and iFlytek’s model is the smallest in size.
In addition, Liu Qingfeng stated that iFlytek Spark X2.6 has率先 broken through the training efficiency bottleneck for long-context tasks on domestic chips, adopting dynamic parallel training technology optimized for domestic platforms to resolve the issue of low hybrid training efficiency caused by limited VRAM bandwidth on domestic chips.
iFlytek has also achieved two significant innovations in the inference efficiency of domestic computing clusters for long-duration tasks:
- One is a distributed inference technology based on domestic heterogeneous computing power, addressing the issue of slow training caused by memory bandwidth limitations of domestic chips.
- Another is to propose development technologies based on domestic equipment, where many things don’t need to be recalculated and can directly use existing results.
Additionally, the Starfire X2.6-4B open-source on-device large model has been officially released, marking a significant leap in on-device large model agent capabilities, with noticeable improvements across all benchmark scores compared to the X2.5-4B model.
IT Home learned from the launch event that since the Spark X2.5-4B edge model was open-sourced on September 1, its downloads have ranked among the top four globally for edge models.
The iFlytek Spark X2.6 series models are now available for trial on the iFlytek Spark official website, featuring the inaugural Spark large model architecture with integrated end-cloud collaboration.
Xunfei Spark X2.6 was trained on domestic computing power, utilizing a MoE architecture with 293B-A30B parameters, an extended context length of up to 1M, and industry-leading code and agent capabilities. It also offers professional security protection and supports customizable security policies tailored to specific task scenarios. The inference pricing is:
- Input is limited to 0.9 yuan per million tokens.
- Input cache hit rate: 0.08 CNY per million tokens
- Output is limited to 3.6 CNY per million tokens.
The iFlytek Spark X2.6-4B open-source on-device large model has significantly improved in code, agent, long-text, and multilingual capabilities. It introduces three new inference modes and achieves higher inference efficiency compared to Ollama under its proprietary inference framework, with inference pricing currently available for free on a limited-time basis.