Most Chinese reasoning models brag about their scores. Spark X1's (星火 X1) more unusual claim is about the factory it was built in: every training step ran on domestic chips, with no foreign accelerators in the loop.
That is a harder engineering feat than it sounds, and it is the real story behind iFlytek's (科大讯飞) January 2025 release.
Key takeaways
- Spark X1 (星火 X1), from iFlytek (科大讯飞), was released on 15 January 2025 as the company's first deep-reasoning large model (大模型) with "slow thinking" capability.
- iFlytek states it is the first deep-reasoning model in China trained entirely on a fully domestic computing platform, using Huawei Ascend (昇腾) 910B accelerators.
- The first release was a 13B model; a 70B version followed on 3 March 2025, according to founder and chairman Liu Qingfeng (刘庆峰).
- The company reports raising Huawei Ascend 910B efficiency from about 20% to nearly 80% of Nvidia-class capability through co-optimization.
- Spark X1 supports a unified fast-and-slow thinking mode and can be privately deployed on as few as four Huawei 910B cards (vendor claim).
Why "domestic" is the headline
iFlytek has been on the U.S. Entity List since 2019, which blocked access to Nvidia and other restricted components. Rather than a workaround, the company turned the constraint into a design principle: train every Spark model on Huawei Ascend (昇腾) silicon. Spark X1 is the clearest expression of that line, described by iFlytek as the only general-purpose deep-reasoning model at its scale trained wholly on a domestic stack.
According to Liu Qingfeng (刘庆峰), chairman of iFlytek (科大讯飞), the 13B version shipped first on 15 January 2025 because compute was constrained, and the 70B version arrived on 3 March 2025 once the domestic cluster was pushed further. The quote captures the constraint honestly: training on domestic chips costs more and takes longer, but the alternative is building "on someone else's foundation."
The efficiency climb
The bottleneck was never the model architecture; it was the chip. Liu stated that through co-optimization with Huawei, the effective efficiency of the Ascend 910B rose from roughly 20% of Nvidia-class performance to close to 80%. That is a vendor-reported figure tied to internal benchmarking, so read it as directional. But the trajectory is the point: a domestic accelerator that starts at a fifth of the efficiency and climbs toward four-fifths changes the economics of sovereign training.
iFlytek built two clusters with Huawei, Feixing-1 (飞星一号) and Feixing-2 (飞星二号), the latter launched in October 2024 as an upgraded large-scale intelligent computing platform. The existence of those clusters is what made X1's fully domestic training possible in the first place.
Reasoning design: fast and slow in one model
Spark X1 is a reasoning model in the "slow thinking" sense: it decomposes hard problems, reflects, and verifies its own steps before answering. iFlytek's notable design choice is a unified model that serves both fast and slow thinking modes, switched by system instruction, rather than shipping separate models. For private deployment this matters: a hospital or school runs one image, not two.
The vendor also claims parity with OpenAI's o1 and DeepSeek R1 on general reasoning. Treat that as a marketing claim, not an audited result; this article did not reproduce those comparisons. Where X1 has the clearest demonstrated strength, per iFlytek's own materials, is in Chinese-language mathematics and in education and healthcare scenarios where the company already has distribution.
Deployment economics
iFlytek states that X1 can be privately deployed on as few as four Huawei 910B cards, with support for both SFT and reinforcement-learning customization. For buyers in regulated sectors (education, healthcare, judiciary) that cannot send data off-site, a compact on-premises footprint is the actual product, more than any leaderboard score. The X1 API is also available on iFlytek's open platform for developers.
A later step in the same line, Spark X1.5 (released November 2025), moved to a Mixture-of-Experts architecture of roughly 293 billion total parameters with about 30 billion active per inference, supporting 130-plus languages. iFlytek describes it as the first MoE model with full-chain training completed on a fully domestic platform. That follow-on extends, rather than replaces, the domestic-training thesis.
Honest limitations
- The "first fully domestic deep-reasoning model" and "80% of Nvidia efficiency" claims come from iFlytek and Liu Qingfeng; they are vendor-stated and not independently verified here.
- The parity claim with OpenAI o1 and DeepSeek R1 is a marketing comparison, not an audited benchmark result in this article.
- Efficiency of domestic accelerators depends heavily on co-optimization; results may not generalize to other labs or other chip batches.
- Spark X1.5's 293B/30B MoE figures are vendor-reported; independent MLPerf-style measurements were not available at writing.
What readers can do now
- If you operate in a regulated sector that requires on-premises inference, evaluate Spark X1's four-card Huawei 910B deployment as a concrete sovereign-compute option.
- Do not accept the o1/R1 parity claim at face value; run your own Chinese-language math and domain tests before deciding.
- Watch the Feixing clusters' published utilization and the X1.5 MoE's active-parameter efficiency as the real signal of whether domestic-only training scales.