iFlytek Spark X1

What it is

iFlytek Spark X1 (讯飞星火 X1) is a large language model built for deep reasoning, released by the Chinese company iFlytek on 15 January 2025 and described as China's first deep-thinking reasoning model trained on a fully domestic computing platform^[1]. A deep-reasoning model is one that spends extra computation working through a problem step by step before answering, instead of producing a reply in a single pass; the approach is aimed at harder mathematics, science and logic tasks.

How it was built

Spark X1 was developed jointly with Huawei^[1]. According to the two companies, the work involved solving problems in how training and inference interact, in high-throughput inference, and in optimising operators — the low-level computational routines — for domestic hardware^[1][2]. That software layer is part of what makes a fully domestic training run difficult: the chips have to exist, and the tooling around them has to work as well.

How it reasons

iFlytek describes the model's reasoning as a process that simplifies a problem first, then explores and verifies candidate solutions reflectively, and says the model is strengthened by training on accuracy-based feedback^[2]. CGTN reported that people familiar with the model described its problem-solving as closer to human "slow thinking"^[2].

Claims and benchmarks

iFlytek said Spark X1 ranked first in China on Chinese-language mathematics, testing it on problems from primary- and secondary-school competitions through university-level ones, as well as on AIME and MATH 500, and that it obtained these results with less compute^[1][2]. The company also said the model used less compute to rank first domestically on several metrics^[1]. Developers said it surpassed OpenAI o1 on some tests, but did not publish the scores behind that claim^[2].

Where it was used

Beyond demonstrations, iFlytek reported deployments in about 100 education pilot areas^[1]. In medicine, it reported 90% accuracy on specialist diagnosis support and on quality control for complex medical records^[1].

Upgrades

A major upgrade announced in April 2025 was claimed to rival OpenAI o1 and DeepSeek-R1 while using a smaller parameter count, and to be trained entirely on Huawei AI chips^[3]. A further upgraded X1 went live on 25 July 2025, benchmarked against OpenAI o3 and supporting more than 130 languages; the X1-0420 version scored over 140 on the 2025 national gaokao mathematics paper I^[4]. The model was succeeded by Spark X1.5 in November 2025^[5].

Where it sits

Several of the central claims around Spark X1 concern domestic computing: the original release was trained on a fully domestic platform^[1], and the April 2025 upgrade was described as trained entirely on Huawei AI chips^[3]. Some comparisons with competing models were announced by the developers without released scores^[2], so the published record of the model consists of company statements alongside the deployments in education and medicine^[1].