Spark X2.5 - iFlytek's flagship general-purpose large model series

What is Spark X2.5

Spark X2.5 is the flagship general-purpose large model series launched by iFlytek, including a cloud-based MoE foundation model (293B/30B) and the open-source on-device models Spark X2.5-4B and Spark X2.5-1.7B. The models are trained through the entire process based on fully domestic computing power, achieving full-stack autonomous controllability. The core upgrades cover code generation, agent collaboration, and mathematical reasoning capabilities. The on-device versions natively support a 1 million Token ultra-long context and adopt a hybrid attention architecture to reduce computational overhead. They are compatible with multiple hardware platforms such as NVIDIA and Huawei Ascend, and support mainstream frameworks such as vLLM,Ollamaand have been made available on the iFlytek Open Platform andHugging Faceand other channels.

Main features of Spark X2.5

• Fully domestic and autonomously controllable: The full training process is completed based on fully domestic computing power, achieving full-stack autonomous controllability from the underlying chips to the upper-layer models.
• Efficient MoE architecture: It adopts a mixture-of-experts architecture with 293B total parameters / 30B activated parameters, balancing strong performance and inference efficiency.
• Intelligent code generation: It greatly improves code understanding and generation capabilities, surpassing larger-parameter models in authoritative code evaluations such as SWE-Bench.
• Autonomous Agent execution: It deeply integrates mainstream Agent frameworks and supports multi-step autonomous planning, tool invocation, and collaborative completion for complex tasks.
• Mathematical logical reasoning: Mathematical and logical reasoning capabilities are significantly enhanced, reaching industry-leading levels in evaluations such as the college entrance examination and AIME.
• Million-Token long context: It natively supports a 1 million Token ultra-long context, the only on-device model with this capability, and can process an entire book or long videos.
• Lightweight open source on device: It open-sources 4B and 1.7B on-device models, adopts a hybrid attention architecture, and runs efficiently on devices such as phones and IoT.
• Full hardware ecosystem compatibility: It seamlessly supports chips such as NVIDIA, Huawei Ascend, Hygon, and Houmo, and is compatible with mainstream frameworks such as vLLM and Ollama.
• Multilingual capability: It supports more than 200 languages, covering major global languages, and meets the needs of cross-language communication and international deployment.

Technical principles of Spark X2.5

• Fully domestic computing power training: The full process of model pre-training and post-training is completed based on domestic intelligent computing clusters such as Huawei Ascend, getting rid of dependence on overseas computing power and achieving full-stack autonomous controllability from the underlying chips to the upper-layer models.
• MoE mixture-of-experts architecture: It adopts a sparsely activated mixture-of-experts design, with 293B total parameters and only 30B activated during inference, significantly reducing inference cost and VRAM usage while maintaining large-model performance.
• Hybrid attention mechanism: The on-device model adopts an architecture alternating 1 layer of full attention and 3 layers of sliding window attention, natively supporting a 1 million Token ultra-long context with extremely low computational overhead, breaking the bottleneck of on-device long-text processing.
• MOPD post-training optimization: Through large-scale reinforcement learning and post-training techniques such as MOPD (Multi-Objective Preference Distillation), it specifically enhances core capabilities such as code generation, autonomous Agent planning, mathematical reasoning, and instruction following.
• On-device lightweight design: The open-source 4B and 1.7B models are deeply optimized for edge scenarios, achieving sub-second response times on low-compute devices such as smart home devices and smartphones, while maintaining task completion quality close to that of cloud-based large models.

Follow on WeChat and reply "open source" to join theAI Open Source Project Exchange Group

How to use Spark X2.5

• API call: Log in to the iFlytek Open Platform orXingchen MaaS platform to obtain a key, and integrate through the standard OpenAI-compatible interface. It is currently free for a limited time.
• Model download: Go to Hugging Face, ModelScope, GitHub, and other channels to download the Spark-X2.5-4B or 1.7B model weight files.
• Local deployment: Choose an inference solution based on your hardware. For NVIDIA GPUs, use vLLM/SGLang; for Huawei Ascend NPUs, pull the official Docker image; for Apple Silicon, use MLX directly.
• One-click run: Import the model through Ollama or LM Studio to start a local chat service without writing any code.
• Custom fine-tuning: Use LLaMA-Factory to load the open-source model for incremental training, quickly adapting it to specific industries or business scenarios.

Core advantages of Spark X2.5

• Fully domestic and independently controllable: The entire training process is completed based on domestic computing power such as Huawei Ascend, achieving full-stack independent control from chips to models and getting rid of external dependence.
• Efficient MoE architecture: With 293B total parameters and only 30B activated, it achieves performance close to that of larger models at lower inference cost.
• Top-tier coding capability: It surpasses 9B-level models in authoritative coding benchmarks such as SWE-Bench, with outstanding algorithm implementation and code completion capabilities.
• Autonomous Agent execution: It is deeply compatible with mainstream Agent frameworks such as Codex and OpenClaw, and supports multi-step autonomous planning and tool invocation for complex tasks.
• Leading mathematical reasoning: Its scores in mathematics evaluations such as the Gaokao and AIME reach the industry's first tier, with significantly enhanced logical reasoning and problem-solving capabilities.
• Million-Token long context: It natively supports ultra-long sequences of 1 million Tokens, the only edge-side model with this capability, and can process an entire book or long videos.
• Lightweight open source for edge devices: The 4B/1.7B models are optimized for devices such as smartphones and IoT, achieving task completion quality close to that of cloud-based large models with extremely small parameters.
• Compatibility with the full hardware ecosystem: It seamlessly adapts to chips from NVIDIA, Huawei Ascend, Hygon, Houmo, and others, and supports one-click deployment with mainstream frameworks such as vLLM and Ollama.

Spark X2.5 project address

• GitHub repository：https://github.com/XHToken/Spark-X2.5
• HuggingFace model library：https://huggingface.co/collections/XHToken/spark-x25

Comparison of Spark X2.5 with similar competitors

Application scenarios of Spark X2.5

• Intelligent programming and development: With top-tier code generation 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 greatly improve development efficiency.
• Enterprise-grade Agent automation: Deeply integrates with mainstream Agent frameworks, supports multi-step autonomous planning, tool invocation, and closed-loop execution for complex tasks, and is suitable for office automation and cross-system business processes.
• On-device long-text processing: The on-device model natively supports an ultra-long context of 1 million tokens, enabling whole-book reading, long contract review, and long-video understanding on offline devices such as phones and in-vehicle systems.
• Intelligent education and math tutoring: It reaches industry-leading levels in math evaluations such as the Gaokao and AIME, and can support personalized math problem-solving tutoring, logical reasoning training, and adaptive learning path planning.
• Fully domestic Xinchuang deployment: Based on fully domestic computing power, it achieves full-stack autonomous and controllable operation, meeting the private deployment needs of industries such as government, finance, and energy that have strict requirements for data security and supply chain autonomy.