iFlytek Unveils Spark X2.6 Trained on Domestic Chips, Upgrading Its Teacher AI Workbench
iFlytek opened its 2026 Global 1024 Developer Festival on October 22 with twin venues in Beijing and Hefei, unveiling Spark X2.6 — a cloud-edge-integrated model the company says is the first to achieve breakthroughs in coding and agent capability while running entirely on domestically produced computing power.
The technical claim goes beyond branding. iFlytek said it trained X2.6 on Chinese-made chips, working around limited on-chip memory and bandwidth with a dynamic parallel training method for long-horizon tasks that lifts ultra-long-context training efficiency by roughly four times. On the inference side, an adaptive compute-memory coordination scheme and what the company calls the first four-layer KVCache built on all-domestic hardware raised inference efficiency by about 4.5 times.
The release also extends to the device. Spark X2.6-4B, trained through a cloud-edge framework, can complete 41.9% of agent tasks locally, and a new routing architecture decides — with 96.1% accuracy, iFlytek said — whether a task stays on-device, moves to the cloud, or is split between the two according to difficulty and data sensitivity.
Education is one of the first domains to receive the upgrade. Xingguang, iFlytek's AI workbench for teachers, lets an instructor describe a need in plain speech — "build me an interactive animation of the magnetic field around a current-carrying solenoid" — and returns classroom-ready interactive courseware that students can drag and manipulate on a classroom display, with optional AR gesture demonstrations of the right-hand rule.
Usage data shared at the event pointed to rapid adoption: satisfaction of 4.7 out of 5 based on multi-dimensional ratings from more than 14,000 classroom teachers, against a reported industry high of 4.0; over 100,000 teacher-created resources per month; and tens of thousands of interactive physics resources authored by teachers on the platform alone.
A second demo covered parent home visits, among the most time-consuming administrative tasks for Chinese teachers. Preparing three visits — identifying priority students, compiling academic and behavioural records, drafting talking points — typically took two to three hours. Xingguang pulls homework data, classroom interaction logs, daily behaviour records and past parent communications into one analysis, cross-checks details (a student dozing in class, for instance, against a parent's note about allergy-related sleep loss) and outputs a preparation form with suggested conversation openers. Teachers cited at the event reported administrative workloads roughly halved and six to ten hours a week returned.
iFlytek attributes the behaviour to three layers: a continuously upgraded domestic foundation model, 22 years of education work with products in more than 60,000 schools, and joint research with institutions including the National Institute of Education Sciences and Beijing Normal University, whose pedagogical frameworks are embedded into expert agents rather than quoted at a model.
The release lands in a competitive context. iFlytek said the previous Spark X2.5-4B on-device model has been downloaded 2.56 million times, placing it among the top four edge models globally, and that this year's developer contest drew 52,120 teams from 351 cities across 15 countries. Chairman Liu Qingfeng framed full-stack self-reliance as training on domestic silicon, not merely inferencing on it.
The company's stated direction for AI in education is to push the technology backstage: teachers should not need to understand models, only articulate what they want to teach, while the assistant absorbs preparation, research, parent communication and routine paperwork.
The technical claim goes beyond branding. iFlytek said it trained X2.6 on Chinese-made chips, working around limited on-chip memory and bandwidth with a dynamic parallel training method for long-horizon tasks that lifts ultra-long-context training efficiency by roughly four times. On the inference side, an adaptive compute-memory coordination scheme and what the company calls the first four-layer KVCache built on all-domestic hardware raised inference efficiency by about 4.5 times.
The release also extends to the device. Spark X2.6-4B, trained through a cloud-edge framework, can complete 41.9% of agent tasks locally, and a new routing architecture decides — with 96.1% accuracy, iFlytek said — whether a task stays on-device, moves to the cloud, or is split between the two according to difficulty and data sensitivity.
Education is one of the first domains to receive the upgrade. Xingguang, iFlytek's AI workbench for teachers, lets an instructor describe a need in plain speech — "build me an interactive animation of the magnetic field around a current-carrying solenoid" — and returns classroom-ready interactive courseware that students can drag and manipulate on a classroom display, with optional AR gesture demonstrations of the right-hand rule.
Usage data shared at the event pointed to rapid adoption: satisfaction of 4.7 out of 5 based on multi-dimensional ratings from more than 14,000 classroom teachers, against a reported industry high of 4.0; over 100,000 teacher-created resources per month; and tens of thousands of interactive physics resources authored by teachers on the platform alone.
A second demo covered parent home visits, among the most time-consuming administrative tasks for Chinese teachers. Preparing three visits — identifying priority students, compiling academic and behavioural records, drafting talking points — typically took two to three hours. Xingguang pulls homework data, classroom interaction logs, daily behaviour records and past parent communications into one analysis, cross-checks details (a student dozing in class, for instance, against a parent's note about allergy-related sleep loss) and outputs a preparation form with suggested conversation openers. Teachers cited at the event reported administrative workloads roughly halved and six to ten hours a week returned.
iFlytek attributes the behaviour to three layers: a continuously upgraded domestic foundation model, 22 years of education work with products in more than 60,000 schools, and joint research with institutions including the National Institute of Education Sciences and Beijing Normal University, whose pedagogical frameworks are embedded into expert agents rather than quoted at a model.
The release lands in a competitive context. iFlytek said the previous Spark X2.5-4B on-device model has been downloaded 2.56 million times, placing it among the top four edge models globally, and that this year's developer contest drew 52,120 teams from 351 cities across 15 countries. Chairman Liu Qingfeng framed full-stack self-reliance as training on domestic silicon, not merely inferencing on it.
The company's stated direction for AI in education is to push the technology backstage: teachers should not need to understand models, only articulate what they want to teach, while the assistant absorbs preparation, research, parent communication and routine paperwork.
JMDedu is a leading B2B EdTech Media in China that informs and connects business professionals through our reports, events, and information services. Our mission is to drive the advancement of education. Our network spans the education industry in China as well as globally, including entrepreneurs, investors, startups, educators, business decision-makers, and policy decision-makers.
Copyright © 2014-2026 en.jmdedu.com All Rights Reserved