At today’s afternoon Global 1024 Developer Festival, iFlytek officially launched the Xunfei Spark X2.6 end-cloud integrated large model. Its most notable feature is that it is the first model based on domestic computing power, which enables the integration of both code and AI capabilities. Chairman Liu Qingfeng emphasized at the press conference that among current large model companies, only Spark uses purely domestic computing power for training. Most others rely on NVIDIA GPUs, and iFlytek’s model size is still the smallest among them.
To improve the efficiency of domestic computing power output, iFlytek has run into the major limitations of domestic chips this time – the memory bandwidth is restricted, and efficiency cannot be improved during long-distance task training. Liu Qingfeng said that X2.6 was the first to break through the bottleneck in training efficiency for long-term tasks using domestic chips. This was achieved through a dynamic parallel training technology tailored for the domestic platform. On the inference side, they introduced two innovations: one is distributed inference technology based on domestic heterogeneous computing power, which addresses the bandwidth bottleneck that causes slow training; the other is development technology based on domestic devices, which allows many intermediate results to be reused directly without having to recalculate everything.
This version on the edge device has also made a step forward. The Xinghuo X2.6-4B open-source edge model is now released simultaneously, marking the first time that an edge model achieves a significant improvement in AI capabilities compared to X2.5-4B. Looking back at September 1st, when X2.5-4B was made open source, its download volume has already ranked it among the top four global edge models. The X2.6 series is currently available for trial on Xunfei Xinghuo’s official website, and the initial release features a collaborative architecture that integrates the edge and cloud models.
The parameters and pricing are also announced. X2.6 is based on domestic computing power training, using a MoE architecture with 293B-A30B parameters, with a context length of 1M. The code and agent capabilities are at the top level in the industry, and it provides professional security protection, supporting custom security policies based on task scenarios. The inference限时 pricing is as follows: 0.9 yuan per million tokens for input, 0.08 yuan for cache hit, and 3.6 yuan for output. And the X2.6-4B sidecar model has seen significant improvements in code, agent, long text, and multilingual capabilities. Three new inference modes have been added, and its efficiency under the self-developed inference framework is better than that of ollama. The inference pricing is free for a limited time.