iFlytek launched the Spark X2.6 end-cloud unified model this Thursday, October 22, 2026, during the Global 1024 Developers Festival. The highlight is being the first to achieve breakthroughs in code and agent capabilities using 100% Chinese national computing, without relying on NVIDIA GPUs.
What happened?
At the event in Beijing, Chairman Liu Qingfeng presented Spark X2.6, based on MoE 293B-A30B architecture with up to 1 million token context. The company claims it is the only mainstream model fully trained on national chips (such as Huawei Ascend), and the smallest in its category. It overcame long-task training bottlenecks on domestic chips with dynamic parallelism techniques.
Simultaneously, the open-source X2.6-4B edge model was released, with significant improvements in agents, code, long text and multilingual over X2.5-4B. X2.5-4B is already among the top four most downloaded edge models globally since September.
Why this matters
Amid advanced chip export restrictions, Spark X2.6 demonstrates advances in Chinese computing autonomy for AI. The end-cloud architecture allows 60% of applications to run locally (more secure and cheaper), 30% with cloud-edge collaboration and 10% in the cloud for complex tasks. This addresses data leakage and token cost concerns.
Inference pricing: input 0.9 yuan per million tokens (cache hit 0.08), output 3.6 yuan. The open-source 4B has limited-time free inference, with three reasoning modes and better efficiency than Ollama on the proprietary framework.
What changes in practice
Users can describe needs in natural language for the model to plan, call tools and generate executable apps. In benchmarks, it competes with or surpasses models like DeepSeek V4.1 Flash in code and agent tasks. The model is available for testing on the official Xinghuo site, with pioneering end-cloud collaborative architecture.
This strengthens the Chinese AI ecosystem in education, health, office and automotive, focusing on security and cost-effectiveness.
By GeekikiBot
