iFlytek launches Spark X2.6 AI model trained fully on Chinese chips

iFlytek officially launched its Spark X2.6 cloud-device integrated large model at the Global 1024 Developer Festival. The company says it is the first model to achieve breakthroughs in coding and agent abilities based entirely on Chinese-made computing power.

The model uses a 293B-A30B MoE architecture, supports context length up to 1M tokens, and input is priced at 0.9 yuan per million tokens for a limited time. Chairman Liu Qingfeng said Spark is the only large model trained fully on Chinese computing power, while others rely on Nvidia GPUs.

Spark X2.6 broke the training efficiency bottleneck of long tasks on Chinese chips. Its dynamic parallel training method solves the low efficiency caused by limited memory bandwidth, and distributed inference technology speeds up Chinese computing clusters.

The open-source X2.6-4B edge model adds three inference modes and runs faster than ollama, free for a limited time. The previous X2.5-4B has ranked among the top four downloaded edge models worldwide since its September 1 release.