iFLYTEK Trains Its Newest Model Entirely on Chinese Chips
iFLYTEK has released Spark X2.6, a model family it says is the first in China to reach a step change in coding and agent work while being trained entirely on domestic computing power. The company announced it at its 1024 developer festival, alongside an open source 4B model for devices and a new speech model called SparkAudio 1.0.
The cloud model is a 293 billion parameter mixture of experts with about 30 billion active per token, and it handles a context window of one million tokens. Chairman Liu Qingfeng said training took the past two to three months.
And the claims rest on benchmarks the company ran itself. iFLYTEK says the model placed first on SWE Pro and SWE Multilingual, second on SWE Verified, and in the top group on enterprise operations agent tasks and the MCP-Atlas tool use test.
The pricing is aggressive, at 0.9 yuan per million input tokens, 0.08 yuan per million cached input tokens and 3.6 yuan per million output tokens, offered as a limited time rate. That is the kind of number that decides whether a developer moves a workload over to test the model, and the cached rate is aimed at the repeat reading that agent loops generate.
Much of the presentation dealt with training without Nvidia. iFLYTEK says its dynamic parallel approach for long sequences works around the memory bandwidth limits of domestic accelerators, and on its Feixing-2 cluster it trained on two trillion tokens of 256K length data in 1.2 months rather than 4.5.
But serving the model matters as much as training it. The company described distributed inference across a mix of domestic hardware and a four tier cache for key value memory built entirely on domestic chips, which is the machinery that decides whether a million token request stays affordable.
The release for local hardware is Spark X2.6-4B, aimed at robots, cars and AI PCs, and iFLYTEK says it beats its own previous 4B model along with Qwen3.5-4B, Gemma4-E4B and Qwen3.5-9B on search, documents, scheduling, spreadsheets and multi step tool calls.
A router sends each task to the device, the cloud or both, and local data is masked before anything leaves the machine. The earlier 4B model is now among the four most downloaded models of its size in the world, according to the company, which points to demand for AI that runs without a network connection.
So the pattern is a large model in the cloud and a small one on the handset, with both trained on Chinese silicon. SparkAudio 1.0 follows the same rule as a speech foundation model trained on domestic compute that combines recognition, translation and semantic understanding with real time speaker separation, sound scene understanding and emotion recognition, already running in the company's input method, its hearing app, its notebooks and its translators.
None of the performance or efficiency figures have been independently verified, and iFLYTEK publishes them itself, so the benchmark wins should be read as company claims until outside testers reproduce them. The models can be tried on the Spark website, and the weights for the 4B version are open.
iFLYTEK released Spark X2.6, a 293 billion parameter mixture of experts with 30 billion active per token.
It has a one million token context and was trained entirely on domestic compute, the company says.
iFLYTEK trained on two trillion tokens in 1.2 months on its Feixing-2 cluster rather than 4.5 months.
Spark X2.6-4B is open source for robots, cars and AI PCs, with a router choosing device or cloud.
All benchmarks and efficiency figures are the company's own.
Source: Pandaily
