iFlytek Launches Spark X2.6 LLM with Enhanced Long-Context and Agent Capabilities

At its 1024 Developer Festival held on October 23, 2026, iFlytek officially unveiled the Spark X2.6 large language model. The new version focuses on long-context understanding and autonomous agent execution, and is designed to run stably on a fully domestic computing infrastructure.

For long-context processing, Spark X2.6 improves information retrieval and memory mechanisms within the context window, reducing the likelihood of forgetting key details, factual errors, and hallucinations when handling lengthy documents such as contracts, technical manuals, and industry files. The agent capabilities have been strengthened with enhanced multi-step task planning, tool invocation, and result validation, enabling the model to break down complex business goals, chain database queries, document read/write operations, and API calls to complete multi-stage tasks. A companion edge-side small model has also been updated to support on-device inference on domestic terminals.

A key highlight is the model's adaptation to domestic computing infrastructure. iFlytek has optimized underlying operators and compilation to bridge models, frameworks, and chips, demonstrating that large general-purpose models can run reliably on domestic hardware. However, challenges remain for large-scale commercialization, including high adaptation costs, performance gaps in cluster computing density compared with overseas high-end hardware, and limited domestic chip supply.