iFlytek Spark model runs on homegrown chips

Chinese technology company iFlytek has released a new iFlytek Spark model that splits work between a device and the cloud, a design the company says brings coding help and autonomous agent tools closer to everyday use. The iFlytek Spark model was unveiled at the company's annual developer festival in Beijing on October 22, where chairperson Liu Qingfeng presented it as a step toward running advanced AI on hardware made in China. The release pairs a large cloud model with a smaller on-device model that developers can use and adapt openly, and the company says the combination keeps sensitive work local while sending harder tasks to remote servers.

The large iFlytek Spark model uses a mixture-of-experts design with 293 billion parameters in total and 30 billion active for each token, and it supports a context window of 1 million tokens, according to reporting on the launch. That window lets the system hold long documents and extended task histories in a single session, which matters for agents that must work through multi-step jobs rather than answer one question at a time. The company said the model was trained and served using computing hardware produced in China, and it described long-horizon training on that hardware as a bottleneck it had worked through with a new parallel training method.

Work splits between device and cloud

The hybrid design of the iFlytek Spark model is the part of the launch most likely to affect developers. A routing system decides whether a task stays on the device, moves to the cloud, or is divided between the two, based on how hard the task is and how sensitive the data looks. The company said the smaller on-device model can finish more than two-fifths of agent tasks on its own, up from about one-eighth in the previous generation. Routing accuracy was reported at over ninety-six percent, the company said, which means most tasks land where the system intended them to land.

For builders, the practical effect is lower cost and less waiting. Simple jobs such as searching documents, drafting schedules, or handling spreadsheet steps can run locally, while the cloud model takes on work that needs broader context or heavier reasoning. The company has also cut the price of calling the iFlytek Spark model through its application programming interface for a limited period, with input priced below one yuan per million tokens and output priced above input. Independent tests of those claims have not yet appeared in English-language coverage, so the performance and pricing figures remain company statements for now, according to a review of the available reporting.

Domestic hardware and everyday uses

The domestic computing claim carries weight beyond engineering. Training on Chinese-made chips speaks to supply chains and data control, two subjects that have shaped AI policy in China as access to foreign hardware has tightened. Liu told the festival audience that the iFlytek Spark model had broken through an efficiency bottleneck in long-horizon training on domestic chips, and an industry analyst quoted in the coverage said the device-cloud combination offers a reference path for the country's AI industry. The emphasis, the analyst said, is shifting from sheer model size toward deployment, data security, and usefulness in offices, factories, and research settings.

In education, a teacher workbench called Xingguang lets an instructor describe a lesson in plain speech and receive interactive courseware that students can drag and manipulate on a classroom display. The example shown at the festival built an animation of the magnetic field around a current-carrying wire, with an optional gesture demonstration of the right-hand rule.

In healthcare, the company showed an imaging cloud for clinical quality control and a system that pulls together complex medical reports to support diagnosis. In cars, a smart cabin system lets a driver state several needs in a single command and receive a route plan shaped by time, budget, and parking. A separate demonstration covered multi-model interaction technology meant to prepare household robots for rooms that change as people move through them.

The festival's developer competition drew teams from hundreds of cities across more than a dozen countries, setting a participation record for the event, according to the company's announcement. Whether the iFlytek Spark model finds users outside China will depend on independent benchmarks, stable pricing after the introductory period, and tools that developers outside the company's home market can readily use. For now, the launch reads as a bet that the next phase of AI competition will be won on deployment and cost, rather than on parameter counts alone. Sources include the company announcement and a People's Daily report on the festival.