26 OCT 2026 — iFLYTEK has released Spark X2.6, a 293-billion-parameter language model that the Chinese company says was trained entirely on domestic computing hardware. It presents the model as its first breakthrough in coding and agent ability on that hardware. Chairman Liu Qingfeng unveiled it on 22 October at the main forum of the company's annual 1024 Global Developer Festival in Beijing, alongside a small open-weights version that runs on the device and a new interaction stack for robots.
The flagship is already on sale through iFLYTEK's API. The 4-billion-parameter sibling, Spark X2.6-4B, is free for now and its weights are downloadable.
What iFLYTEK launched
According to the company's Spark API product page, Spark X2.6 is a mixture-of-experts model with 293 billion total parameters, 30 billion of them active, and a context window of 1 million tokens. The page says it was trained on domestic compute and describes its coding and agent abilities as "industry first-class" (our translation), a phrase that is iFLYTEK's own assessment. The festival's official programme lists the main forum for 14:00 to 15:30 on 22 October at Phase II of the China National Convention Center.
Prices on the same page are marked as a limited-time 40% discount: 0.9 yuan per million input tokens, 0.08 yuan per million cached input tokens and 3.6 yuan per million output tokens. The previous model, Spark X2.5, is listed at 0.96 yuan in and 3.6 yuan out, so the newer model is slightly cheaper on input at today's promotional rates. The page does not say when the discount ends.
The domestic-chip claim
Liu framed the launch around self-reliance. "Only by truly training your own model on domestic compute is it really autonomous and controllable," he said, according to an account of the speech carried by The Paper (our translation). IT Home reported him telling the audience that Spark was the only model among its peers trained wholly on domestic compute, with the others trained on Nvidia cards. RECATOOLS has not checked that comparison, and iFLYTEK named no rivals.
The company says it worked around the memory and bandwidth limits of Chinese chips with a dynamic parallel training method for long tasks, which it credits with roughly four times faster training on very long contexts. On the inference side it describes a four-layer KV cache built on domestic hardware and claims a 4.5-fold efficiency gain. Those are vendor figures. The accounts we read do not name the training chips or publish the measurements behind them, and we have not seen an independent benchmark of X2.6.
A small model that routes work
The more practical idea for developers may be the device-cloud split. iFLYTEK calls it a device-cloud architecture: the 4B model running locally decides whether a task stays on the machine, goes to the cloud model, or is shared between the two. The company says its router picks correctly 96.1% of the time. In a stage demo, the small model reviewed a purchase contract by masking sensitive details locally, sending the masked text to the cloud for checking after the user agreed, and then restoring the details on the device.
The model card for Spark X2.6-4B, published on Hugging Face under the name SparkLLM, says the model keeps exactly the same architecture as X2.5-4B and gets its gains from better post-training. It offers three thinking levels (none, low and high) and is released under the Apache 2.0 licence. The card claims state-of-the-art results among open models of similar size, "matching or even surpassing models with approximately 9B parameters". Again, that is the publisher's own benchmarking.
Robots that talk while they move
Robotics got its own segment. Liu argued that robots leaving sports competitions and stage shows for daily life must hear instructions as distance and noise change, follow several people at once and coordinate speech with movement. "Human-computer interaction has always pulled industrial revolutions along," he said, in the same account (our translation).
iFLYTEK's answer is Spark embodied multimodal interaction technology. The company lists four advances: noise reduction in changing spaces, working out which of several people intends to interact, planning speech and action together, and starting an interaction when it sees a visual cue. A demo showed a robot in a mock supermarket fetching goods while answering a family of four whose requests kept changing. The technology is packaged in Robot Super Brain Platform V3.0, which bundles a Spark Super Brain board, an embodied-intelligence developer kit, a multimodal interaction module and an embodied interaction harness.
Tools for coders and businesses
For software teams, iFLYTEK says X2.6 works with its agent process-automation tooling and an upgraded command-line tool to support conversational development, connections to enterprise systems and tool calls. The festival also brought a speech foundation model, SparkAudio 1.0, and a platform for AI digital employees called WinWork that the company says already offers more than 80 types of digital worker. According to Science and Technology Daily, the launches also included AI PC products.
What is still unknown
Spark X2.5-4B, the previous small model, has been downloaded 2.56 million times, according to iFLYTEK, which says that puts it among the top four on-device models worldwide. X2.6-4B will be judged on the same terms. Its card shows benchmark charts as images, without third-party reproductions, and the accounts we read give no task list behind the 41.9% figure for on-device agent tasks. Developers comparing it with other 4B-class models will want their own tests first. A profile of the earlier models is in our AI directory entry for iFLYTEK Spark.
View author profile → · Editorial policy
