iFlytek (Spark) review

Yes. A free quota is offered for Spark API applications across the models in Spark Lite, Spark Pro and Spark Pro 128K, so an early build can run before any paid commitment. The amount of that quota is not set out here, so size a pilot against the current terms rather than a fixed allowance.

The catalogue is Spark Lite, Spark Pro and Spark Pro 128K. That covers Spark Lite, Spark Pro, Spark Pro-128K, Spark Max, Spark Max-32K and Spark 4.0 Ultra, alongside deep reasoning and literature models.

The largest window is 131,072. Model names in Spark Lite, Spark Pro and Spark Pro 128K hint at the same order of magnitude for the long-context tiers, which matters for whole-document work.

Spark sells API access with a free quota for new applications, and Spark Lite, Spark Pro and Spark Pro 128K is the catalogue you would price tier by tier. The listing carries no single blended rate, so cost comparison means checking each model rather than a single headline number.

Spark comes from iFlytek, founded in 1999, and the models are the company's own rather than a resold set. That matters when tracking a vendor release: new tiers such as Spark 4.0 Ultra in Spark Lite, Spark Pro and Spark Pro 128K arrive from the same builder that runs the API.

Support runs on chat, help centre and community, which covers live chat, a help centre and a community forum. That is enough for setup questions and API errors, though a phone or email route is not part of the listed channels, so plan ahead on anything urgent.

The Spark models are documented by the vendor itself, and Spark Lite, Spark Pro and Spark Pro 128K lists every tier it publishes. No regional breakdown is given for the API, so check the terms that apply in your market before you commit a production workload to it.

Spark keeps a project inside Spark Lite, Spark Pro and Spark Pro 128K with one vendor relationship. A gateway such as Supascale AI or Azure AI Foundry offers a far wider catalogue, so the trade-off is catalogue breadth against keeping the whole stack with the company that builds the models.