iFlytek Releases Spark X2.6 Trained on Domestic Compute; X2.5-4B Hits 2.56 Million Downloads
At the Beijing leg of its 1024 Developer Festival on October 22, iFlytek released Spark X2.6, a large model trained entirely on domestic compute via the Feixing platform built with Huawei on the Ascend ecosystem and capable of supporting trillion-parameter training. Chairman Liu Qingfeng said an autonomous compute base 'offers the world a second choice,' and through dynamic parallel training iFlytek lifted ultra-long-context training efficiency fourfold and inference efficiency 4.5-fold on Feixing-2 chips, raising KV-cache hit rate from 63% to 95.7%. The end-side Spark X2.5-4B model surpassed 2.56 million downloads a month after open-sourcing, entering the global top four end-side models, and end-side agent task coverage rose from 12.8% to over 40%.
Source: China Jilin Network, Oct 22, 2026 | 2026-10-22
- Zhou Hongyi Says DeepSeek's Open Source Reflects 'Sovereign AI'; 360 Builds Sovereign-AI OS
- AgiBot's Expedition A3 Sets Guinness Record With 110 Humanoid Robots Dancing in Unison
- iFlytek Releases Spark X2.6 Trained on Domestic Compute; X2.5-4B Hits 2.56 Million Downloads
- Galbot and Unitree Jointly Develop World's First Native-Agent Robot 'GalbotET1'
- Don't Be Misled by Token Consumption, Says Economic Daily: China-U.S. Top Model Gap Narrows Below 3%
- AI Is Getting Things Done, Accelerating the 'Compute-Model-Scenario' Ecosystem Chain
- iFlytek Debuts Xingchen TaaS Platform at 1024 Developer Festival; MiniMax Open-Sources M3.1 Flash
- DeepSeek and MiniMax Complete AI Model Overseas-Registration, Clearing Compliance Path
- 2026 Yiwu Trade Conference Closes, Spotlighting AI as Engine of Cross-Border Commerce
- Yiwu's Small Commodities 'Borrow' Hainan's Chengmai to Reach the Globe
