iFlytek pairs Spark X2.6 with a token-as-a-service platform that bundles models, agents and industry solutions
The AMW Read
An incremental update from an established Chinese model lab: a new Spark version, an open 4B edge model and an agent/TaaS bundle. The impact stays mostly within China's enterprise LLM market.
iFlytek pairs Spark X2.6 with a token-as-a-service platform that bundles models, agents and industry solutions
At its 1024 Global Developer Festival in Beijing on October 22, iFlytek chairman Liu Qingfeng launched Spark X2.6 with a 1M-token context window and open-sourced Spark X2.6-4B, an on-device model. A new device-cloud routing setup, trained with a critique model, raised routing accuracy from 81.2% to 96.1%. The company also released SparkAudio 1.0, a speech foundation model trained entirely on domestic Chinese compute. iFlytek reports 20% better accuracy, 15% lower latency and 60% lower call cost for it. The open platform added Agent OS for coordinating agents across enterprise systems, APA for turning screen recordings or plain-language instructions into reusable skills, an XFYUN CLI that reaches 700+ AI capabilities, and Xingchen TaaS, which sits on its MaaS layer and serves both Spark and open-source models.
The TaaS pitch is the strategic signal. iFlytek is moving away from selling tokens and model calls and toward selling packaged workflows for finance, healthcare, education and energy customers. That's a defense against falling per-token prices from Chinese open-weight rivals. Its on-device open-source line also seems to be working: iFlytek says Spark X2.5-4B has been downloaded 2.56 million times since we covered its September release. The company now claims 11.87 million developers on its platform. All of these figures come from iFlytek itself. The article is a sponsored reprint, and none of the numbers are independently verified.
For builders, the new piece to watch is the device-cloud routing. Hardware makers can run a free 4B model locally and send harder queries to paid Spark endpoints in the cloud. That pushes adoption of the open model toward cloud lock-in. Investors should check whether TaaS revenue shows up separately from model API income in future disclosures.


