iFlytek releases Spark X2

Background

iFlytek's Spark line of large models began with V1.0 in 2023 and went through several iterations. V3.0 was positioned around Chinese-language capability, and with V4.0 the company compared the model's overall performance with GPT-4 Turbo^[2]. Each step in that sequence raised the same practical question: where the computing power behind training comes from, and whether a model can keep improving on a stable supply of it.

The release

On 11 February 2026 iFlytek announced Spark X2 through its official Weibo account^[1]. The company described the release as a comprehensive upgrade of general capabilities^[1], and stressed that X2 had been trained on fully domestic computing power^[1].

"Domestic computing power" here means the chips, servers and computing platform used in training came from suppliers inside the country rather than imported accelerators. For a model developer this matters for supply: training a large model takes large numbers of accelerators over long periods, and a domestic stack means the work does not depend on access to hardware from abroad.

Where it is aimed

Spark X2 was presented as targeting education, healthcare, automotive and agent scenarios^[1]. An agent, in this context, is a model set up to call tools and carry out a task in several steps rather than only answering a single question — a form of use that has become a common target for model makers.

Healthcare received the most specific description. iFlytek said X2 was tuned for diagnostic accuracy and treatment recommendation, and for parsing complex medical records^[2]. Medical records are largely unstructured text — doctors' notes, test results, histories written in inconsistent formats — so handling them well is a different problem from answering a general question, and one that depends heavily on domain tuning.

Market position

Coverage of the release noted that iFlytek ranked first in large-model tender wins in 2025, with contracts worth RMB 2.316 billion^[2]. Tenders are the public procurement process through which buyers — government bodies, hospitals, schools, enterprises — solicit bids and select a supplier, so the figure is a measure of how much large-model work the company was contracted to do in that year rather than a measure of model quality.

What it means

Spark X2 sits at the intersection of two themes in the Chinese large-model field. One is the shift from general capability benchmarks toward named industry scenarios, where what counts as good performance is defined by a customer's task — reading a medical record correctly, for instance — rather than by a general test. The other is the compute stack: announcing that a model was trained entirely on domestic computing power makes that supply chain part of the product claim, not just an engineering detail^[1].

Against the earlier generations, the company's framing is continuity rather than a break: V1.0 in 2023, V3.0 oriented to Chinese, V4.0 compared with GPT-4 Turbo, and now X2 described as an across-the-board upgrade on a domestic stack^[1][2]. The release therefore says as much about the infrastructure the company intends to train on as about the model itself.