机器之心
1 sourcesiFlytek Releases Spark X2.5: Trained on 10,000 Domestic GPUs, AI Feeds Back into Compute Optimization
On September 7, iFlytek officially released the Spark X2.5 large model, featuring a MoE architecture with 293B-A30B parameters, enhanced code and agent capabilities, and support for over 200 languages. The model was pre-trained and post-trained on a 10,000-GPU domestic 910B cluster, with online inference on domestic platforms and multi-step speculative decoding. Notably, Spark X2.5 has begun participating in NPU operator optimization, allowing AI to feed back into underlying compute optimization and providing practical insights for engineering challenges in domestic large-scale training, such as operator efficiency, long context, communication, and stability.