- The 293-billion-parameter model is trained and deployed entirely on Chinese computing infrastructure
- Its smaller open-source versions bring million-token context to phones, PCs and robots
Chinese AI company iFlytek has launched its latest flagship model, Spark X2.5, with a focus on coding and agentic capabilities as AI moves from answering questions to completing tasks.
Released September 7, the model uses a mixture-of-experts (MoE) architecture with 293 billion total parameters and 30 billion activated parameters.
iFlytek says it was trained and deployed entirely on domestic computing infrastructure and is now available through its open platform.
The Yangtzeer reported earlier that the Hefei-based company had released two smaller open-source versions, Spark X2.5-4B and 1.7B, on September 1.
Designed for phones, PCs and robots, they are the first edge models to natively support context windows of up to 1 million tokens, according to iFlytek.
That gives local devices the ability to handle long documents, assist with coding and perform other tasks that have traditionally relied on cloud-based models.
From conversation to action
Coding and agentic AI are at the center of the X2.5 upgrade.
The 4B model is roughly one-tenth the size of some cloud models, yet iFlytek says it can match models two to three times larger on tasks including algorithm implementation and code completion.
The flagship model also improves tool calling, task planning and autonomous execution. Given a set of raw data, for example, it can analyze the information, generate a report and translate it into multiple languages without requiring the user to direct every step.
Why it matters globally
For global AI developers, the launch highlights two trends worth watching.
The first is the push to make smaller models increasingly capable on edge devices. A million-token context window on a 4B model could make local AI more useful for professional and industrial applications while reducing dependence on the cloud.
The second is the growing use of domestic computing infrastructure in China’s AI stack.
X2.5 does not by itself prove that Chinese hardware has closed the gap with Nvidia, but training and deploying a flagship model entirely on domestic platforms offers another sign that China is building greater independence across the AI computing chain.
Header image generated by Doubao
