iFlytek Spark X2.6

What it is

Spark X2.6 is a model family from the Chinese AI company iFlytek, built to run partly on a user's device and partly in the cloud instead of entirely in a data centre. The cloud side is a mixture-of-experts (MoE) architecture with 293 billion parameters^[3]. In a mixture-of-experts model, a routing layer sends each piece of input to a small subset of specialised sub-networks rather than running the whole model every time, so the computation spent on a request is much smaller than the headline parameter count would suggest. The family also includes smaller models that run locally on phones, PCs, tablets, AI glasses, car cockpits and robots^[1].

How it was announced

iFlytek chairman Liu Qingfeng unveiled the family on 22 October 2026 at the 2026 iFlytek Global 1024 Developer Festival in Beijing^[1][2]. The company said the models were trained on fully domestic computing power^[1][2].

The router that decides where a task runs

The central mechanism is a routing model that looks at each request and chooses one of three execution modes: on-device only, device-cloud collaboration, or cloud only^[1]. The choice takes into account how difficult the task is, what the local device can handle, and whether the data involved is sensitive^[1]. Simple tasks can therefore be handled locally on a phone, PC, tablet, AI glasses, smart cockpit or robot, while harder ones are passed to the cloud^[1]. iFlytek puts the accuracy of the routing decision at 96.1%^[1].

Data handling

Because some of the work stays local, the family is designed so that sensitive data can be masked on the device and sent to the cloud only after the user confirms it, and the original local files are not uploaded at all^[1].

Efficiency on domestic hardware

iFlytek reports that long-context training efficiency improved by about 4x and inference efficiency by 4.5x^[1]. The gains came from using a four-layer KVCache on domestic hardware together with distributed inference^[1]. A KVCache stores intermediate attention state computed for earlier tokens so that it can be reused for later ones instead of being recomputed; stacking several layers of it and spreading the work across machines is one way to make long inputs cheaper to process.

Code and agent capability

iFlytek described the release as its first breakthrough in code and agent capability achieved on fully domestic compute^[1]. The model can plan the steps of a task, call external tools, and generate runnable applications from a request written in ordinary language^[1]. On the device side, the X2.6-4B model was trained with an adaptive post-training method and reached 41.9% coverage of agent tasks^[1].

How it is offered

The family is available as on-device model downloads, as a cloud API, and inside iFlytek products including Spark AStudio, Spark Loomy and the Spark app^[1].

Where it sits

The device-cloud design is a response to a trade-off that model makers meet as assistants move onto phones, cars and robots: sending everything to the cloud gives access to the strongest model but costs latency, bandwidth and privacy, while running everything locally limits what the model can do. Spark X2.6 handles that trade-off by making the placement of each task an explicit decision — keep sensitive data on the device, and give the difficult work to the cloud^[1].