Spark X2-Flash - A large language model based on the MoE architecture launched by iFlytek.

What is Spark X2-Flash?

Xinghuo X2-Flash is a large language model based on the MoE architecture released by iFlytek. It has a total of 30 bytes of parameters, supports 256KB of ultra-long context, and is trained on a domestic computing cluster powered by Huawei Ascend 910B. Designed specifically for the Agent era, the model performs close to trillion-parameter models in scenarios such as agent task execution, code generation, and deep research, while its token consumption cost is less than one-third of mainstream large-scale models. The model achieves efficient training and inference through techniques such as DSA sparse attention and MTP multi-token prediction. Its API is open and integrated with platforms such as AstronClaw and Loomy.

Main functions of Xinghuo X2-Flash

• Intelligent agent task executionIt supports complex agent workflows such as in-depth research report generation, skill management and invocation, and system control and execution, with results approaching those of a trillion-parameter model.
• Code generationIt can quickly generate complex skills (such as AI video-generated skills), including complete descriptions of skill structure, core functions, and use cases.
• Long context processingIt supports a maximum context window of 256K, which can meet the consumption needs of hundreds of thousands or even millions of tokens in long-chain Agent tasks.
• Multi-platform accessIt has been integrated with products such as AstronClaw and Loomy, and is compatible with mainstream agent frameworks such as OpenClaw and Claude Code.
• API serviceThe Xingchen Coding Plan fully supports this model by providing API calls through the iFlytek Open Platform and Xingchen MaaS Platform.

Technical Principles of Spark X2-Flash

• MoE architectureThe model employs a hybrid expert architecture with a total of 30 parameters, achieving higher efficiency while maintaining performance.
• Domestic computing power trainingTraining was completed based on the Huawei Ascend 910B cluster, and deep optimization was achieved through operators that are compatible with domestic chips and distributed training strategies.
• Intelligent agent data closed loop: Construct a verifiable large-scale intelligent agent data automatic synthesis platform, in which the agent autonomously builds the environment and tests the accuracy of the results to achieve efficient data synthesis and closed-loop.
• Efficient Training of Long TextsIt is the first domestic computing power to combine DSA (sparse attention) and MTP (multi-token prediction), with the context extended to 256K, and the training efficiency increased from 20% to 90% compared with the A800 cluster of the same size.
• Sampling decoding efficiency optimizationIn reinforcement learning training scenarios, through algorithmic and engineering innovations, sampling and decoding efficiency can be improved by more than 2 times, alleviating the computational barrier for RL training in long-interaction scenarios.

Key information and usage requirements of Spark X2-Flash

• Model NameSpark X2-Flash
• PublisheriFlytek / iFlytek Open Platform
• Model ArchitectureMoE (Hybrid Expert), Total Parameters 30B
• Context windowMaximum support 256K
• Training computing powerHuawei Ascend 910B domestic cluster
• Already connected to the platformAstronClaw, Loomy
• API EntryiFlytek Open Platform, Xingchen MaaS Platform
• Compatible frameworkMainstream agent frameworks such as OpenClaw and Claude Code
• Usage requirements:
  • Developers can call the API through the iFlytek Open Platform or the StarMaaS Platform.
  • The Starry Sky Coding Plan fully supports this model, and both new and existing users can switch to it independently.

The core advantages of Spark X2-Flash

• Extremely high cost performanceThe performance of complex agent tasks is close to that of models with trillions of parameters, while token consumption is less than one-third of that of mainstream large-scale models.
• Domestic computing power is independent and controllableTraining is based on Huawei Ascend 910B clusters, and it runs efficiently on local computing power architecture.
• Extremely long contextA 256K context window meets the long-link requirements of complex agent workflows.
• Breakthrough in training efficiencyThrough DSA+MTP technology, the training efficiency of domestically produced computing power has been increased from 20% to 90%.
• Fast reasoning speedSampling and decoding efficiency is improved by more than 2 times, and the training time for reinforcement learning is significantly reduced.
• Agent native optimizationDeeply compatible with mainstream agent frameworks such as OpenClaw, supporting automatic closed-loop data synthesis for intelligent agents.
• Rapid Ecosystem AccessIt has been integrated with applications such as AstronClaw and Loomy, and developers can use it immediately.

Comparison of Spark X2-Flash with similar competing products

Application scenarios of Spark X2-Flash

• Complex Agent Workflow: In-depth research report generation, multi-step task breakdown and execution, and multi-round context reading and correction.
• Skill/Tool DevelopmentAutomatically generate and manage complex skills (such as AI video generation skills), including structure definition, core functions, and use cases.
• Code generation and system controlScenarios requiring coding skills, such as script writing, system command execution, and automated operation and maintenance.
• Long document analysisIt can process extremely long documents, papers, and reports based on a 256K context, and perform abstracting, extraction, and question answering.
• Multimodal task orchestrationAs the brain of the Agent, it coordinates multiple platform toolchains such as text-based video and image-based video (e.g., Keling, Runway, Pika).