Spark X2.5 Operational

Spark X2.5 is iFlytek's flagship general-purpose large model series, featuring cloud-based MoE base models (293B/30B) and open-source edge models Spark X2.5-4B and Spark X2.5-1.7B. The models are trained through an end-to-end process using fully domestic computing power, achieving full-stack autonomy and controllability. Core upgrades enhance code generation, agent collaboration, and mathematical reasoning. The edge versions natively support an ultra-long context of 1 million Tokens and use a hybrid attention architecture to reduce computational overhead. They are compatible with multiple hardware platforms, including NVIDIA and Huawei Ascend, and support mainstream frameworks such as vLLM and Ollama. They are available through channels including the iFlytek Open Platform and Hugging Face.

Fully domestic and independently controllable: The entire training process is completed using fully domestic computing power, achieving full-stack autonomy and controllability from underlying chips to upper-layer models.
Efficient MoE architecture: Uses a mixture-of-experts architecture with 293B total parameters and 30B activated parameters, balancing strong performance with inference efficiency.
Intelligent code generation: Significantly improves code understanding and generation, surpassing larger-parameter models on authoritative coding benchmarks such as SWE-Bench.
Autonomous agent execution: Deeply integrates with mainstream agent frameworks and supports multi-step autonomous planning, tool use, and collaborative completion of complex tasks.
Mathematical and logical reasoning: Mathematical and logical reasoning capabilities are significantly enhanced, reaching top-tier industry performance on evaluations such as the Gaokao and AIME.
1-million-Token long context: Natively supports an ultra-long context of 1 million Tokens, the only edge model with this capability, enabling the processing of entire books or long videos.
Lightweight open-source edge models: Open-sources 4B and 1.7B edge models that use a hybrid attention architecture to run efficiently on devices such as phones and IoT equipment.
Broad hardware ecosystem compatibility: Seamlessly supports chips from NVIDIA, Huawei Ascend, Hygon, Moore Threads, and others, and is compatible with mainstream frameworks such as vLLM and Ollama.
Multilingual capabilities: Supports more than 200 languages, covering major languages worldwide and meeting the needs of cross-language communication and international deployment.

Follow us on WeChat and reply “开源” to join the AI open-source project discussion group
Training on fully domestic computing power: The full pre-training and post-training process is completed using domestic intelligent computing clusters such as Huawei Ascend, reducing reliance on overseas computing power and achieving full-stack autonomy and controllability from underlying chips to upper-layer models.
MoE mixture-of-experts architecture: Uses a sparsely activated mixture-of-experts design with 293B total parameters, of which only 30B are activated during inference, significantly reducing inference costs and GPU memory usage while maintaining large-model performance.
Hybrid attention mechanism: The edge models use an architecture alternating one full-attention layer with three sliding-window attention layers. This natively supports a 1-million-Token ultra-long context with extremely low computational overhead, overcoming the bottleneck of long-text processing on edge devices.
MOPD post-training optimization: Uses large-scale reinforcement learning and post-training techniques such as MOPD (multi-objective preference distillation) to specifically enhance core capabilities including code generation, autonomous agent planning, mathematical reasoning, and instruction following.
Lightweight edge design: The open-source 4B and 1.7B models are deeply optimized for edge scenarios, delivering sub-second responses on low-computing-power devices such as smart home appliances and phones while maintaining task-completion quality close to that of cloud-based large models.

API access: Log in to the iFlytek Open Platform or Xingchen MaaS platform to obtain an API key, then connect through the standard OpenAI-compatible interface. Currently free for a limited time.
Model download: Download the Spark-X2.5-4B or 1.7B model weights from channels such as Hugging Face, ModelScope, and GitHub.
Local deployment: Choose an inference solution based on your hardware. Use vLLM/SGLang with NVIDIA GPUs, pull the official Docker image for Huawei Ascend NPUs, or use MLX directly on Apple Silicon.
One-click execution: Import the model through Ollama or LM Studio to start a local chat service without writing code.
Custom fine-tuning: Load the open-source model with LLaMA-Factory for incremental training and quickly adapt it to specific industries or business scenarios.

Fully domestic and independently controllable: The entire training process is completed using domestic computing power such as Huawei Ascend, achieving full-stack autonomy and controllability from chips to models and reducing external dependence.
Efficient MoE architecture: Activates only 30B of its 293B total parameters, delivering performance close to that of larger models at a lower inference cost.
Leading coding capabilities: Surpasses 9B-level models on authoritative coding benchmarks such as SWE-Bench, with outstanding algorithm implementation and code completion capabilities.
Autonomous agent execution: Deeply compatible with mainstream agent frameworks such as Codex and OpenClaw, supporting multi-step autonomous planning and tool use for complex tasks.
Advanced mathematical reasoning: Achieves top-tier industry results on mathematical evaluations such as the Gaokao and AIME, with significantly enhanced logical deduction and problem-solving capabilities.
1-million-Token long context: Natively supports an ultra-long sequence of 1 million Tokens. It is the only edge-model product with this capability and can process entire books or long videos.
Lightweight open-source edge models: The 4B/1.7B models are optimized for devices such as phones and IoT equipment, delivering task-completion quality close to that of cloud-based large models with very few parameters.
Broad hardware ecosystem compatibility: Seamlessly adapts to chips from NVIDIA, Huawei Ascend, Hygon, Moore Threads, and others, and supports one-click deployment with mainstream frameworks such as vLLM and Ollama.

GitHub repository: https://github.com/XHToken/Spark-X2.5
Hugging Face model collection: https://huggingface.co/collections/XHToken/spark-x25

Intelligent software development: With leading code-generation capabilities and SWE-Bench performance, it supports the full software engineering process from requirements analysis to testing and deployment, and can serve as an AI programmer to significantly improve development efficiency.
Enterprise-grade agent automation: Deeply integrates with mainstream agent frameworks and supports multi-step autonomous planning, tool use, and closed-loop execution for complex tasks, making it suitable for office automation and cross-system business processes.
Edge long-text processing: The edge models natively support a 1-million-Token ultra-long context, enabling offline devices such as phones and in-vehicle systems to read entire books, review lengthy contracts, and understand long videos.
Intelligent education and mathematics tutoring: Reaching top-tier industry performance on mathematical evaluations such as the Gaokao and AIME, it can support personalized math tutoring, logical reasoning training, and adaptive learning-path planning.
Fully domestic deployment for information technology innovation: Full-stack autonomy and controllability based on fully domestic computing power meets the private-deployment needs of industries such as government, finance, and energy that have strict requirements for data security and supply-chain autonomy.