iFlytek launches Spark X2.6, claiming training entirely on domestic computing

iFlytek launched its Spark X2.6 model family on October 22, 2026, promising stronger coding and AI agent capabilities from training entirely on domestic computing resources. Trial access and an open-source on-device model are available; the claims about overcoming hardware limits and improving performance remain iFlytek’s account of its own work.

The company introduced the family at its Global 1024 Developer Festival. IT Home’s launch account, republished by KuCoin, describes an integrated cloud-and-device offering: Spark X2.6 is available for trial through the official Spark website, while X2.6-4B is the separately released open-source model intended to run on devices. The two models belong to the same release, but carry different specifications and performance claims.

The domestic-computing claim reaches into training

Chairman Liu Qingfeng put domestic computing at the center of the launch. He asserted that Spark was the only model trained on domestic computing resources, contrasting it with models trained on NVIDIA graphics processors. The more specific technical case concerns how iFlytek says it handled long-context training on domestic chips.

iFlytek attributes its training gains to dynamic parallel training optimized for domestic platforms. It says the approach addresses low training efficiency caused by limited video-memory bandwidth—the rate at which data moves through that memory. For inference, or running the trained model, it describes distributing work across different domestic computing resources and reusing existing results rather than recalculating them.

A long-context model with published usage rates

Spark X2.6 is described as a mixture-of-experts model, using selected parts of the model for computation, with a listed architecture of 293B-A30B. Its stated context limit is up to one million tokens, the units of text a model processes. That figure describes the amount of material it can work with in a request, rather than a score for coding or task completion.

The published pricing separates ordinary input, cached input and generated output. Cached input has its own lower rate; output has the highest of the three listed rates.

CNY 0.9Input

The listed rate for input supplied to Spark X2.6.

CNY 0.08Cached input

The separate listed rate for cached input.

CNY 3.6Output

The listed rate for text generated by the model.

The on-device release has its own performance pitch

The open-source X2.6-4B companion targets on-device use. iFlytek says it improves coding, agent work, long-text handling and multilingual capabilities over X2.5-4B, with higher scores across the benchmarks cited at the launch.

iFlytek also says X2.6-4B introduces three inference modes and runs more efficiently than Ollama under its proprietary inference framework.