At the recent iFlytek 1024 Global Developers Festival, chairman Liu Qingfeng released Spark X2.6 — a 293-billion-parameter MoE flagship whose complete training remained on domestic compute, Observer reported on October 28. It is the first time a million-token ultra-long-context code and agent training run was completed on domestic chips, with overall capability benchmarked against the global first tier.
A simultaneously released X2.6-4B open-source on-device model, trained via an end-cloud architecture, strengthens long-text understanding, multi-turn tool calling and cross-tool coordination, covering 41.9% of agent tasks completable on-device. While peers put parameters and leaderboards at center stage, iFlytek prioritized a harder path: full training and inference must run on a fully domestic chain.
Liu framed full-stack autonomy as four layers — domestic-platform inference and training adaptation, self-developed algorithms, autonomously obtained and cleaned data, and scenario software-hardware integration — with training the deepest "blue ocean." The "15th Five-Year Plan" and the State Council's "AI+" opinion both embed "secure and controllable" in their requirements.
The progress signals that domestic compute platforms, through sustained algorithmic and engineering innovation, can keep iterating model capability rather than merely serving inference.
Source: Observer / iFlytek
