Overview
At the iFLYTEK 1024 Global Developer Festival on Oct 22, iFLYTEK released Spark X2.6, an end-cloud integrated LLM that achieves breakthroughs in coding and agentic capability on fully domestic computing. The model uses a 293-A30B MoE design (293B total, 30B active per token), supports a 1-million-token context, and runs on a domestic 10,000-GPU inference cluster (Feixing-2).
Training uses dynamic parallel training for long sequences to work around domestic accelerators’ memory-bandwidth limits: on Feixing-2, training 2,000B tokens of 256K-length data took 1.2 months versus 4.5, lifting long-context training efficiency ~4x. Inference deploys adaptive compute-storage co-inference, heterogeneous distributed inference and a four-tier KV-Cache on domestic hardware, improving inference efficiency 4.5x with KV-Cache hit rate rising from 63% to 95.7%.
Background & Interpretation
1. Technology Background and Evolution
With tightening U.S. AI-chip export controls, coupling domestic models with domestic compute is inevitable. Spark X2.6 stresses full-stack autonomy—not just inference but training on domestic platforms. Chairman Liu Qingfeng calls end-cloud integration the answer to AI’s two thresholds: fear of data leakage and fear of high cost.
2. Core Drivers and Underlying Mechanism
End-cloud splits work by complexity: simple tasks stay on phones, PCs, glasses, cockpits and robots; complex tasks go to the cloud. This saves tokens and protects privacy via on-device masking. iFLYTEK also open-sourced Spark X2.6-4B for robots, vehicles and AIPC; its X2.5-4B reached 2.56M downloads, among the world’s top four on-device models.
3. Market Structure and Industry Impact
The launch shifts China’s LLM race from “parameters and benchmarks” to “domestic-compute fit, data security and real deployment.” Cloud pricing of 0.9 yuan per million input tokens targets government, SOE, education, healthcare and judicial scenarios. Deeper trend: closer model-chip co-design (e.g., Huawei Ascend) moving China’s AI ecosystem from “usable” to “practical.”
Implications & Outlook
Near term, domestic-compute LLM breakthroughs accelerate on-prem deployment in data-sensitive sectors. Medium term, end-cloud and on-device open-source lower AI barriers, spawning desktop agents and embodied interaction. Watch domestic-cluster cost curves, on-device ecosystems and export-control evolution.
Takeaways for Industry Participants
- Government and enterprise users should evaluate “domestic compute + end-cloud” for balanced security and cost.
- Developers should watch open on-device models for lightweight innovation.
- Chip and model vendors must deepen software-hardware co-design to turn hardware constraints into system advantages.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.