iFLYTEK introduced the Spark X2.6 family of device-cloud models at its Global Developer Festival 1024 on October 22, 2026. The company claims this is the first model in China to achieve significant progress in coding and agent capabilities while being fully trained using internal computing power.
Chairman Liu Qingfeng stated that training the model took the last two to three months. The cloud model utilizes a 293B-A30B Mixture-of-Experts architecture, meaning a total of 293 billion parameters, with approximately 30 billion active per token. It supports a context window of one million tokens.
According to iFLYTEK's own tests, the model ranked first in the SWE Pro and SWE Multilingual benchmarks, and second in SWE Verified. Furthermore, it entered the leading group for corporate operations agent tasks and the MCP-Atlas tool usage test. The temporary pricing is 0.9 yuan per million input tokens, 0.08 yuan per million cached input tokens, and 3.6 yuan per million output tokens.
A significant part of the presentation was dedicated to training and inference on domestic chips. iFLYTEK explained that dynamic parallel training for long sequence tasks allows it to overcome memory bandwidth limitations of domestic accelerators. On the Feixing-2 cluster, training on 2 trillion tokens of 256K length data took 1.2 months instead of the initially planned 4.5 months.
For inference, the company described distributed inference on heterogeneous domestic hardware and a four-level KV Cache architecture built entirely on in-house equipment. Additionally, the company released the source code for the Spark X2.6-4B model, designed for robots, automobiles, and AI PCs.
iFLYTEK stated that Spark X2.6-4B surpasses its predecessor Spark X2.5-4B, as well as models Qwen3.5-4B, Gemma4-E4B, and Qwen3.5-9B in areas such as search, document processing, planning, spreadsheets, and multi-step tool calling. The previous 4B model, which was released on September 1st, is now among the four most downloaded device models globally, according to the company.
The routing system determines whether a task will be executed on the device, in the cloud, or distributed between them, with local data being masked before being sent to the cloud. iFLYTEK also introduced SparkAudio 1.0—a foundational speech model trained on domestic computing power. This model integrates speech recognition, translation, and semantic understanding with real-time speaker diarization, sound scene understanding, and emotion recognition. It is already integrated into iFLYTEK Input Method 16.0, iFLYTEK Hearing, its smart notebooks, and translators.
All presented performance metrics and benchmarks are the property of iFLYTEK. The cloud and device models are available for testing on the Spark website.