Spark X1

Unveiled January 15, 2025, iFlytek's Spark X1 is a deep-reasoning large language model notable less for raw scale than for how it was built: trained entirely on domestic computing infrastructure — 10,000 Huawei Ascend 910B cards — as a deliberate demonstration of China's self-reliant AI stack. It emulates human slow thinking, breaking problems down with self-reflection and verification loops, and iFlytek reported it matching DeepSeek R1 (671B) and OpenAI o1 on math benchmarks while leading on Chinese math tasks at a ~70B reference scale.

A March 2025 upgrade (Spark X1) extended the line into verticals: the first Spark medical model built on X1, plus iterations of the AI learning machine and teacher-assistant products. A further generation, Spark X1.5, followed in November 2025 with MoE training-efficiency improvements and 130+ language support.

Limitations: Spark X1 is accessed through iFlytek's products and platform, not as open weights — there is no self-hosting path for independent developers. Most performance claims are vendor-reported against domestic benchmarks; independent English-language evaluation is thin, and international availability is limited.

Vendor-reported: on par with DeepSeek R1 and OpenAI o1 in math; leads Chinese math tasks