TL;DR
iFLYTEK released the on-device-cloud large model Spark X2.6 on October 22, 2026. It uses a 293B-A30B MoE architecture trained entirely on domestic compute with 1M-token context; training efficiency rose about 4x and inference efficiency 4.5x, and the on-device model covers 41.9% of agent tasks.
Background: The 1024 Developer Festival and a Domestic Compute Base
On October 22, 2026, the main forum of the 2026 iFLYTEK Global 1024 Developer Festival opened at the China National Convention Center in Beijing, where iFLYTEK Chairman Liu Qingfeng unveiled the on-device-cloud Spark X2.6. The event showcased the Feixing Platform, built by iFLYTEK together with Huawei on the Ascend ecosystem; iFLYTEK describes it as the country's first deployed domestic 10,000-GPU compute base, capable of supporting trillion-parameter model training.
The Model: Spark-X2.6 Architecture and Capabilities
According to the official Spark model page, Spark-X2.6 uses a 293B-A30B MoE (mixture-of-experts) architecture, was trained on domestic compute, and extends context length to 1M tokens, with coding and agent capabilities benchmarked against leading global models. iFLYTEK says this marks the first time it has achieved breakthroughs in code generation and long-horizon agent tasks on entirely domestic compute, and that the model provides professional security protection with task-scenario-based safety policies.
Training and Inference Engineering
To work around the memory and bandwidth limits of domestic chips, iFLYTEK disclosed two engineering advances. On the training side, a dynamic parallel training technique for long-horizon tasks adaptively allocates compute by task length, improving long-context training efficiency by about 4x. On the inference side, adaptive compute-memory collaborative inference on domestic clusters, combined with a four-layer KV Cache technique built entirely on domestic hardware and distributed inference, improves inference efficiency by 4.5x.
On-Device-Cloud: X2.6-4B and the Collaboration Architecture
Spark X2.6 strengthens both cloud and on-device capabilities and supports three execution modes: on-device only, on-device-cloud collaboration, and cloud only. The on-device model decides the path based on task difficulty, its own capability, and data-security requirements. iFLYTEK says Spark X2.6-4B, trained with an on-device-cloud scheme, strengthens long-text understanding, multi-turn tool use, and cross-tool coordination, reaching 41.9% agent-task coverage on-device; at the Qingdao city summit, the company disclosed that on-device-cloud routing accuracy rose from 81.2% to 96.1%. The approach also lets sensitive information be desensitized on the device, reducing data leaving the device.
Product Matrix and Industry Applications
Alongside Spark X2.6, iFLYTEK released the SparkAudio 1.0 speech foundation model, embodied multimodal interaction technology, the Xingfan WinWork digital-employee platform, and the iFLYTEK AIPC series. iFLYTEK says WinWork has deployed more than 80 types of digital employees covering risk control, contract review, and sourcing, and that relying on the on-device-cloud Spark model saves 44% of token consumption. The on-device-cloud model is being applied first in education, healthcare, and legal services.
Sources
- Science and Technology Daily (China Science and Technology Network, October 23, 2026): http://www.stdaily.com/web/gdxw/2026-10/23/content_601627.html
- Guangming Online (October 23, 2026): https://digital.gmw.cn/2026-10/23/content_39062959.htm
- Spark official model page: https://xinghuo.xfyun.cn/sparkapi?ch=aigc-lhy
- Qingdao News Network citing Guanhai News (October 24, 2026): http://m.toutiao.com/group/7700246243229860403/
- Global People Magazine (Toutiao, October 24, 2026): http://m.toutiao.com/group/7700239929984549395/