iFlytek's New Model: Edge-Cloud in One, on Domestic Compute
iFlytek launched Spark X2.6 at its 1024 Global Developer Festival on October 22, positioned as an edge-cloud integrated model: a 293-billion-parameter mixture-of-experts model in the cloud with a one-million-token context window, and a 4B on-device model open-sourced at the same time. The company says the model was trained entirely on domestic compute.
- 01THE LAUNCH — On October 22, iFlytek launched Spark X2.6 at the 1024 Developer Festival.
- 02THE CLOUD MODEL — A 293B-A30B mixture-of-experts model in the cloud, with a one-million-token context window.
- 03THE DEVICE MODEL — The 4B on-device model is open-sourced, with edge-cloud routing accuracy at 96.1%.
WHAT HAPPENEDWhat Happened
The cloud model's architecture is a mixture-of-experts design with 293 billion total parameters and 30 billion activated per token, and a context window of one million tokens. The company says it ranks first on SWE Pro and SWE Multilingual and second on SWE Verified, that training time was compressed from 4.5 months to 1.2 months, and that inference efficiency rose 4.5 times.
The device side is a separate line: the open-source Spark X2.6-4B on-device model and an edge-cloud collaborative architecture were shown for the first time, with task-routing accuracy up from 81.2% to 96.1% and the share of tasks completed by the device alone up from 12.8% to more than 40%. Pricing is equally aggressive: after a limited-time discount, 0.9 yuan for input and 3.6 yuan for output per million tokens.
THE COMPUTEDomestic Compute, From Slogan to Training Ledger
The company says all of Spark X2.6's training was completed on domestic compute, with its Feixing-2 ten-thousand-card cluster running a two-trillion-token training load in 1.2 months that previously took 4.5. The figures come from the vendor and need third-party reproduction, but the direction is clear: the domestic compute bottleneck is shifting from whether training is possible to how fast it can be done.
The supporting engineering detail matters too: long-sequence dynamic parallelism, distributed inference and four-level key-value caching are designs aimed at compensating for domestic accelerators' memory-bandwidth shortcomings. Half of a model launch event is, in effect, a compute engineering event.
THE ROUTINGEdge-Cloud Coordination Is a More Concrete Selling Point Than Parameters
The edge-cloud architecture means something specific: simple tasks stay on the device, complex tasks go to the cloud, and a routing model decides the split. The routing-accuracy rise from 81.2% to 96.1% directly determines user experience and cloud cost; parameter scale takes a back seat.
Open-sourcing the 4B on-device model amplifies the strategy: device makers can take a free local model and route heavy tasks to the paid cloud. The previous generation of the on-device model accumulated 2.56 million downloads, showing the distribution path has already worked once.
THE TAASFrom Selling Tokens to Selling Packaged Workflows
Released alongside the model were Agent OS, a recorded-workflow skills tool and Xingchen TaaS. TaaS packages the model, agents and industry solutions into deliverable workflows rather than charging by the token.
The positioning answers the price war: as token prices for domestic open-source models keep sliding, vendors need revenue beyond tokens. Whether TaaS becomes a revenue line of its own depends on whether future financial reports disclose it separately.
THE SHIFTWhat Changed
SIGNAL JUDGMENT — ANALYSISParameters Are Not the News; the Compute Ledger Is
The Spark X2.6 launch covered two things: a model, and an engineering ledger for domestic compute. Tokenta Signal's judgment is unchanged: the next phase of competition among domestic models will be decided not by whose parameter table is longer, but by who can drive down training and inference costs on their own compute.
WHAT REMAINS UNCERTAINWhat Remains Uncertain
EVIDENCEEvidence
The original report confirms the launch points: the October 22 1024 Developer Festival, the edge-cloud Spark X2.6, the open-source 4B on-device model, and 52,120 teams taking part in the developer competition.
PRIMARY SOURCESPrimary Sources
- EQS / Media OutReach: iFLYTEK Unveils Spark-X2.6 and a Range of New Products
https://www.webdisclosure.com/press-release/iflytek-etr-iflytek-unveils-spark-x26-and-a-range-of-new-products-4s5FEph7iLW - Intelligent Living: iFLYTEK Spark X2.6: China's Homegrown AI for Coding and Robots
https://www.intelligentliving.co/iflytek-spark-x2-6-china-homegrown-ai/ - AI Market Watch: iFlytek pairs Spark X2.6 with a token-as-a-service platform
https://www.ai-market-watch.com/news/iflytek-launches-spark-x26-sparkaudio-10-and-its-agent-os-platform-uc7y9f
RELATEDRelated Intelligence
- Update: Apple Intelligence Beta Goes Live in China, With Qwen Inside Siri
- Nvidia Reportedly Tells Cloud Partners to Cap Chinese Customers at 49% of New Contracts
- China Accelerates Full-Lifecycle AI Legislation: DeepSeek, Moonshot AI and Alibaba in Scope
- OpenAI and Anthropic set up election war rooms for the first time: the midterm election day duty test