This IT Home page is stamped 2026/10/22 14:40:44. iFlytek released Spark X2.6 that afternoon at its global 1024 developer festival. The page calls it the first code and agent step achieved with training on domestic compute only. Chairman Liu Qingfeng says Spark is the only one trained on all-domestic compute, that the others are all trained on N cards, and that iFlytek's size is the smallest. Those are his words at the event. This page attaches no training records from the other labs, and no parameter comparison table.
[1]
The page says Spark X2.6 was trained on domestic compute, uses a 293B-A30B MoE, and extends context length to 1M. Calling the code and agent ability first-rate in the industry is this page's wording, not a rank on a public board. The series is open for trial on the Spark site, and the page calls the release the first on-device and cloud architecture of its kind here. Input is a limited 0.9 yuan per million tokens, a cache hit is a limited 0.08 yuan, and output is a limited 3.6 yuan. Each price says limited time. No end date is printed.
[1]The same page releases the open on-device model Spark X2.6-4B. It calls that model a step up for on-device agents, and says every score is clearly higher than X2.5-4B. The scores themselves are not printed. What opened on September 1 was the previous X2.5-4B. The page says that since that model opened, downloads have entered the global top four for on-device models. That is the previous model's downloads, not a new figure for X2.6-4B. X2.6-4B is described as much stronger on code, agents, long text, and multiple languages, and as adding three inference modes. The page does not name the three modes or say how they differ. It says efficiency under its own inference framework is higher than ollama. It does not say by how much. Inference pricing for this on-device model is free for a limited time, and that end date is not printed either.
[1]Liu also says X2.6 is first to break the training-efficiency bottleneck for long tasks on domestic chips. The page also names two claims about inference efficiency. This story keeps the fact that those claims were made, and does not write the training or inference procedure. A line about safety protection and task-specific safety policy is kept as a line on the page, not as instructions for setting a policy. A reader who wants a scoreboard of X2.6 against the previous version will not find one here. The input, cache, and output prices are printed on this page. They are not an estimate.
[1]