iFlytek Spark X2.6 was officially released on October 22, 2026, at the main forum of the 2026 iFlytek Global 1024 Developer Festival in Beijing. It is an edge-cloud integrated large model, and iFlytek says it is the first to achieve a breakthrough in code and agent capabilities trained entirely on domestic Chinese compute: the cloud version uses a mixture-of-experts architecture with 293B total parameters and 30B active parameters, extends context length to 1M, and is already available for trial on the Spark website; the Spark X2.6-4B on-device model released at the same event is open-sourced outright.
Domestic compute: the bottlenecks were training and inference
Memory capacity and bandwidth limits on domestic chips have historically hurt long-horizon tasks most. iFlytek offered two fixes this time: on the training side, a dynamic parallel training technique for long tasks allocates compute adaptively by task length, raising ultra-long-context training efficiency by about 4x; on the inference side, adaptive compute-memory co-scheduling, the first four-layer KV cache built on domestic hardware, and distributed inference together deliver an officially claimed overall efficiency gain of about 4.5x. Chairman Liu Qingfeng put it more bluntly on stage: Spark is trained on fully domestic compute, at a relatively smaller model size. That is the vendor's own framing, and third-party benchmarks will be the real test, but at least the training and inference efficiency numbers are now on the table.
How edge and cloud split the work, and where pricing lands
Liu's proposed split: about 60% of applications finish locally, with data never leaving the device and no inference cost; about 30% run through edge-cloud collaboration; the remaining 10% go to the cloud model alone. Cloud pricing is a limited-time input rate of 0.9 yuan per million tokens, 0.08 yuan per million for cache hits, and 3.6 yuan per million for output, with security policies customizable per task scenario. The on-device X2.6-4B improves markedly over its predecessor on code, agents, long text and multilingual ability, adds three inference modes, and is free for inference for a limited time; its predecessor, X2.5-4B, has ranked among the top four on-device models globally by downloads since its open-source release on September 1, 2026.
What makes this launch worth watching is not the spec sheet itself, but that iFlytek turned whether domestic compute can train a competitive model into a public answer: training efficiency, inference caching, a small on-device model and pricing all landed at once. For government and enterprise buyers with domestic-substitution requirements, there is now one more option they can benchmark themselves; for developers, a free on-device 4B and a cheap cloud API mean they can run real tasks first and decide afterwards whether to bet on this route.
