On October 22, iFlytek released the cloud-edge integrated Spark X2.6 model family at its 1024 Developer Festival, alongside the SparkAudio 1.0 speech foundation model, a full-modal in-cabin interaction system and the WinWork digital-employee platform. The company says Spark X2.6 is the first to achieve code and agent capability breakthroughs trained fully on domestic compute.

Tech illustration of a smartphone linked to cloud servers by light beams
Cloud-edge integration: sensitive data stays local, heavy tasks go to the cloud

1. What "trained on domestic compute" really means

Chairman Liu Qingfeng said Spark is currently the only model in the industry trained entirely on domestic AI chips — others train on Nvidia GPUs and at most migrate afterward to domestic clusters. Amid chip-supply uncertainty, the hardest path has become supply-chain certainty. According to Guandian.cn, Spark X2.6's long-context training efficiency improved about fourfold.

2. Cloud-edge integration: a 4B model on device

Spark X2.6 achieves an agent-capability leap on the edge, with a new cloud-edge architecture: an on-device 4B model handles sensitive information locally, and coordination efficiency between device and cloud models reaches 96.1%. For users, voice and document data can stay on the device; for enterprises, edge inference cuts cost and latency.

3. Aggressive API pricing and a full product stack

Pricing is set at a limited-time 0.9 RMB per million input tokens, 0.08 RMB for cache hits, and 3.6 RMB per million output tokens — among the lowest tier of domestic APIs. The multimodal cockpit supports multi-passenger simultaneous interaction with 95% accuracy on complex travel intents, and a new recording gadget transcribes 2-hour recordings in 2 minutes.

Spark X2.6 at a glance

· Training base: fully domestic compute (claimed industry-unique)

· Cloud-edge coordination efficiency: 96.1%; on-device model: 4B

· Limited-time API: RMB 0.9/M input tokens, RMB 3.6/M output tokens

As the cloud-side model arms race plateaus, competition is shifting to edge deployment and industry depth — whoever gets models into phones, cars and office terminals wins the next round. Cloud-edge collaboration is also a deployment pattern of NineZenith's Tianxing models: sensitive data processed locally, complex tasks orchestrated in the cloud.

(Information synthesized from Leikeji, Wallstreetcn, Beijing News and other public reports)