iFlytek launches Spark X2.6 large model trained entirely on domestic compute

The headline claim here is less the model than the substrate: a flagship general-purpose system trained end-to-end on domestic compute, which is the constraint that has bound Chinese labs most tightly. The code and long-horizon agent gains matter because those are the workloads enterprises pay for, and they are where domestic stacks have lagged. The open question is whether full-domestic training scales beyond a single release into repeatable iteration cycles — and whether the reported agent-task breakthroughs hold up under independent evaluation rather than festival-stage benchmarks.