iFlytek Launches Spark X2.5, a Flagship AI Model Trained Entirely on Domestic Chinese Silicon

Chinese artificial intelligence labs are under pressure to prove that frontier-grade models can be built without access to the most advanced American chips, and iFlytek used a launch on September 7 to make that case in public. The Hefei-based company, best known for its voice recognition technology and for supplying AI tools to the Chinese government and state-owned enterprises, announced the official release of Spark X2.5, a flagship large language model with a mixture-of-experts architecture. The headline detail was not the benchmark score but the supply chain: iFlytek said the model was trained and is being served entirely on domestic computing infrastructure, with no dependence on overseas GPUs.

Key Facts

iFlytek unveiled Spark X2.5 on September 7, according to the Chinese technology publication ITHome, which reported that the model uses a mixture-of-experts design with 293 billion total parameters and roughly 30 billion active parameters per token, a configuration written as 293B-A30B. The same report said the model supports a 256K context window and covers more than 200 languages, positioning it for both Chinese-language tasks and international deployment.

ScienceNet, the online news service of the Chinese Academy of Sciences, reported on September 7 that the release puts a new emphasis on code generation and agentic capabilities, the two areas where Chinese labs have been racing to close the gap with American frontier models. Science and Technology Daily, the state-backed newspaper, added on the same day that Spark X2.5 completed its entire training and inference lifecycle on a domestically produced computing platform, a claim that matters politically as well as technically because it signals that Chinese AI leaders can operate without Nvidia's most advanced accelerators.

iFlytek also published pricing that is aggressive even by Chinese standards. ITHome reported on September 7 that access to Spark X2.5 costs 1.6 yuan per million input tokens, 0.24 yuan per million tokens on a cache hit, and 6 yuan per million output tokens, roughly a fraction of the price of equivalent Western frontier APIs after currency conversion. The model is available on iFlytek's own MaaS platform and its open platform for developers.

The launch follows an earlier release of smaller models. Science and Technology Daily reported on September 7 that iFlytek had already open-sourced Spark X2.5's edge-side variants at 4B and 1.7B parameters, which support context windows of up to one million tokens for on-device use cases such as document processing, coding assistance and data analysis.

The two-tier release strategy mirrors what several Chinese labs have adopted this year: keep the largest model on a paid API, give away small models as open weights, and let developers distribute the brand. iFlytek's version of that playbook leans on its unusual positioning among Chinese AI companies. It is not a pure research lab like the startups that dominate the benchmark leaderboards, and it is not a cloud giant like Alibaba or Huawei. It is a listed company with a decades-old voice and education business, state-linked customers, and a hardware ecosystem that includes smart speakers, translators, recording pens and classroom devices, all of which become deployment targets for a model that can run on domestic chips.

Analysis

What this really means is that iFlytek is betting its corporate future on the argument that China no longer needs to import the most advanced AI hardware to compete, and Spark X2.5 is the most credible evidence it has offered to date. A 293B-parameter MoE model trained end to end on domestic silicon is not the same as GPT-6 class scale, but it is large enough to be genuinely useful, and the fact that iFlytek chose to run the full stack on domestic chips rather than quietly renting overseas capacity is a strategic statement to Beijing, to its enterprise customers and to its own shareholders.

The bigger picture here is that the export-control regime has split the Chinese AI market into two tracks. One track, exemplified by labs that can still obtain advanced accelerators through various channels, keeps training at the frontier. The other track, which now includes iFlytek, is optimizing for a world in which the marginal GPU is a domestic chip, and the competitive metric shifts from raw scale to efficiency, pricing and integration with Chinese cloud and government infrastructure. Spark X2.5 will not unseat the global frontier models on pure capability, but it does not need to; its moat is that it runs on infrastructure that American export rules cannot reach, and that it is priced and tuned for the Chinese enterprise and public-sector market.

The pricing deserves attention. At 1.6 yuan per million input tokens, iFlytek is undercutting most Western APIs by an order of magnitude, and the cache-hit price of 0.24 yuan per million tokens is designed to win high-volume, repetitive workloads such as customer service automation, where iFlytek already has deep distribution through its legacy voice business. The strategy reads as a classic platform play: monetize the flagship to fund the ecosystem, open-source the small models to seed developer mindshare, and lean on the state's preference for domestic technology to win anchor deployments.

Why It Matters

For Chinese enterprises and government agencies, Spark X2.5 offers a politically safe, reasonably capable and very cheap alternative to models that depend on foreign chips, and that combination will matter far more than leaderboard position. For iFlytek, the stakes are existential: the company's growth has lagged the pure-model labs, and this release is its attempt to convert its distribution advantage in education, healthcare and public services into a genuine foundation-model business with recurring API revenue.

For the global AI industry, the significance is structural. A credible flagship model trained entirely on domestic Chinese hardware narrows the assumption that frontier AI is permanently tied to American semiconductor supply, and it gives other countries under export pressure a reference point for building their own stacks. It also intensifies price competition in the API market, since Chinese labs have repeatedly shown they will compress margins to gain share, and Spark X2.5's pricing continues that pattern.

The support for more than 200 languages points to a quieter ambition that is easy to overlook. If Spark X2.5 performs well in Southeast Asian, Middle Eastern, African and Latin American languages, iFlytek can position it as a sovereign-friendly alternative to American models for governments and carriers that do not want their data flowing through US-controlled infrastructure, an advantage that no amount of parameter scaling by a foreign lab can fully neutralize. That is the same playbook that has made China the leading supplier of telecom equipment to the developing world, applied to software.

Next Up

The most useful signals in the coming months will be adoption rather than announcement. Watch whether Spark X2.5 wins large government and state-owned enterprise contracts, whether iFlytek publishes third-party benchmark results that hold up under scrutiny, and whether the company can convert its open-source edge models into an on-device ecosystem that reaches developers outside China. The longer-term question is whether the domestic silicon that trained this model can scale to the next generation, and whether iFlytek's next release closes the capability gap with the labs that still train on unrestricted American hardware. A useful date to watch is the next quarterly earnings call, where iFlytek will have to show whether API revenue from Spark X2.5 is growing fast enough to justify the investment.