StarryFire X2.5 – The Flagship General-Purpose Large Model Series from iFLYTEK

1. What is StarryFire X2.5

StarryFire X2.5 is iFLYTEK's flagship general-purpose large model series, which includes the cloud-side MoE base model (293B/30B) and open-source edge-side models StarryFire X2.5-4B and StarryFire X2.5-1.7B. This series is fully trained using entirely domestically produced computing power, achieving full-stack autonomy and controllability from the underlying chips to the upper-layer models. Its core capabilities focus on three major directions: code generation, intelligent agent collaboration, and mathematical reasoning. The edge-side version natively supports ultra-long context with over 1 million Tokens, and employs a hybrid attention architecture to significantly reduce computational overhead. The model is compatible with multiple hardware platforms such as NVIDIA and Huawei Ascend, supports mainstream inference frameworks like vLLM and Ollama, and has been officially released through the iFLYTEK Open Platform and Hugging Face, among other channels.

Technical Positioning and Domain: StarryFire X2.5 belongs to the domain of general-purpose large language models, covering two major application directions: high-performance cloud-side inference and lightweight edge-side deployment. Its unique positioning lies in its ability to cover all scenarios—from data centers to edge devices—through a single series that includes a MoE cloud-side model with a total of 293B parameters and edge-side models with 1.7B and 4B parameters. It is one of the few product lines among domestic large models that achieve a coordinated layout between the cloud and edge.

Development Background: This series was primarily developed by the iFLYTEK Research Institute, leveraging iFLYTEK's long-term technical accumulation in areas such as voice interaction, educational AI, and industry applications. The development motivation directly targets two key pain points of domestic large models: computing power autonomy and efficient edge-side inference. By deeply integrating with domestic intelligent computing clusters such as Huawei Ascend, it explores a full-stack training path independent of overseas computing power, providing a compliant option for industries with high security requirements such as government and finance.

Core Value: StarryFire X2.5 addresses two critical issues in domestic large models: supply chain security and the ability to process long texts on the edge. Technologically, it is the first edge-side model to natively support a context length of over 1 million Tokens, breaking through the bottleneck of processing ultra-long sequences on mobile devices. On the industrial level, it provides a reference model for full-chain autonomy from chips to applications, reducing the risk of dependence on imported computing power for key industries.

2. Key Features

• Fully Domestic and Self-Controllable Training: Conducts the entire pre-training and post-training process using domestic intelligent computing clusters such as Huawei Ascend, without relying on any overseas computing resources. This feature gives StarFire X2.5 a natural advantage in supply chain security and data compliance, making it particularly suitable for industries such as government, finance, and energy that have strict restrictions on the source of computing power.

• MoE Efficient Architecture Design: The cloud-based base model employs a Mixture-of-Experts (MoE) architecture with a total of 293B parameters and 30B activated parameters. By leveraging a sparse activation mechanism, it significantly reduces the computational load during inference. Compared to similarly scaled dense models, this design maintains generation quality while substantially reducing memory usage and the cost per inference, balancing performance with economic efficiency.

• Code Intelligence Generation and Understanding: Demonstrates code capabilities surpassing those of larger parameter models in authoritative code benchmarks such as SWE-Bench, covering scenarios like code completion, algorithm implementation, bug fixing, and unit test generation. iFLYTEK claims its 4B edge-side model achieved a score of 44.4 on the SWE-Bench Pro benchmark, reaching a leading level among small models, and can support real-world software engineering tasks.

• Agent Autonomous Planning and Execution: Deeply integrated with mainstream Agent frameworks such as Codex and OpenClaw, it supports multi-step autonomous planning, tool calling, and result verification for complex tasks. The model is capable of breaking down task objectives, calling external APIs, integrating multi-source information, and forming a closed-loop execution, making it applicable to automated workflows and cross-system business collaboration scenarios.

• Enhanced Mathematical and Logical Reasoning: Mathematical reasoning capabilities are enhanced through MOPD post-training technology, achieving industry-leading performance in evaluations such as Gaokao (Chinese college entrance exam) mathematics and AIME. The model can perform multi-step symbolic derivation, inequality proof, and complex application problem solving, supporting intelligent education and research assistance scenarios.

• Million-Token Ultra-Long Context Support: Natively supports a context window of up to 1 million Tokens, making it the only edge-side model with this capability. Users can directly input entire books, lengthy technical documents, or several hours of video content without the need for slicing, greatly expanding the information processing scope of a single interaction.

• Lightweight Open-Source Edge Models: Open-sources two edge-side versions with 4B and 1.7B parameters, employing a hybrid architecture that alternates between one layer of full attention and three layers of sliding window attention. This enables sub-second response times on low-computing-power devices such as smartphones, smart home appliances, and in-vehicle systems, while maintaining task completion quality comparable to large cloud models.

• Full Hardware Ecosystem Compatibility: Seamlessly supports mainstream chips both domestically and internationally, including NVIDIA, Huawei Ascend, Hygon, and PostMo, and is compatible with inference frameworks such as vLLM, SGLang, Ollama, and MLX. Developers can flexibly choose deployment options based on their existing hardware resources.

• Multilingual Coverage Capability: Supports text generation and understanding in over 200 languages, covering major global languages. This capability enables StarFire X2.5 to be applied in cross-language scenarios such as cross-border e-commerce content generation, multilingual customer service, and international news translation, reducing language barriers in global deployment.

3. How to Use

1. Environment and Prerequisites: To call the cloud API, register an account on the iFLYTEK Open Platform or the Star MaaS Platform and apply for an API key. For local deployment, prepare an NVIDIA GPU with at least 16GB of VRAM (recommended RTX 4090 or A100) or a Huawei Ascend NPU environment. The model can run on mobile devices with a minimum of 4GB of memory.

2. API Integration Process: After logging into the iFLYTEK Open Platform and obtaining an API key, call the cloud model through an OpenAI-compatible interface. The request format is consistent with the OpenAI standard interface. Existing projects can be migrated by simply modifying the base_url and api_key. The official currently provides a limited-time free API call quota.

3. Model Weight Acquisition: Download the Spark-X2.5-4B or Spark-X2.5-1.7B model weight files from Hugging Face, ModelScope, or GitHub. After downloading, load the weights using the Hugging Face Transformers library or vLLM to perform inference and testing in the local environment.

4. Local Inference Deployment: For NVIDIA GPU environments, it is recommended to use vLLM or SGLang for high-throughput inference. For Huawei Ascend NPU, pull the official Docker image and configure the CANN toolchain. Apple Silicon devices can directly run the model using the MLX framework, without requiring additional adaptation.

5. Zero-code Desktop Execution: Import the model weights using Ollama or LM Studio. After completing the graphical interface configuration, you can launch a local chat service. This method is suitable for non-technical users to quickly experience the model without writing code or handling complex dependencies.

6. Custom Fine-tuning: Use the LLaMA-Factory framework to load open-source model weights. After preparing a domain-specific dataset, perform LoRA or full-parameter incremental training. This allows for adaptation and optimization tailored to specific industry terminology, writing styles, or business logic within a few hours.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Scenario Selection Recommendations: For industries such as government, finance, and energy that have strict requirements for self-reliance in the computing power supply chain, StarFire X2.5 is the preferred choice due to its fully domestic training and multi-hardware compatibility, especially in scenarios requiring on-device offline processing of ultra-long documents. For internet and software development teams seeking top-tier code capabilities and multimodal applications, Qwen3.7-Max performs more outstandingly on code benchmarks such as SWE-Bench, but its closed-source strategy and binding to the Alibaba Cloud ecosystem must be carefully considered during selection. For academic institutions and small to medium-sized enterprises requiring deep customization and private deployment, DeepSeek V4 provides greater flexibility with its MIT open-source license and adaptable deployment options, although its context length is significantly shorter compared to StarFire X2.5.

6. Editor's Summary

The technological innovation of StarFire X2.5 is primarily reflected in three aspects: first, the verification of a full-process training pathway using domestic computing power, completing the entire chain from pre-training to post-training on the Huawei Ascend cluster, establishing a replicable engineering paradigm for domestic large models to break free from reliance on overseas computing resources; second, the architectural breakthrough in on-device models supporting native 1 million Token context, achieving the ability to process ultra-long sequences under parameter scale constraints through an alternating layout of full attention and sliding window attention. This design approach provides a valuable reference for addressing the computing power bottlenecks in on-device AI; third, the MOPD multi-objective preference distillation post-training technology integrates three core capabilities—code, mathematics, and Agent—into a unified optimization framework, offering greater advantages in training efficiency and capability balance compared to traditional single-task reinforcement learning.

In terms of practical value, the open-source strategy for the on-device 4B and 1.7B models effectively lowers the entry barrier for developers. The Apache 2.0 license removes legal obstacles for commercial integration, while full hardware ecosystem compatibility allows existing computing devices to be reused. This series is suitable for key industry users with information innovation compliance requirements, software developers needing to process long documents on mobile devices, and small and medium-sized enterprise technical teams looking to access Agent capabilities at a low cost. However, due to the lack of open-sourcing for cloud-based models and the absence of multi-modal capabilities, its competitiveness in scenarios requiring deep customization and multi-modal integration still has room for improvement.

7. Application Scenarios

• Intelligent Programming and Software Development: Developers can leverage the code generation capabilities of StarFire X2.5 to assist with daily coding tasks, including function implementation, bug fixing, unit test generation, and code review. Its performance on the SWE-Bench evaluation indicates its ability to handle certain automated programming tasks, and it can be deployed as a code review assistant within CI/CD pipelines, significantly improving team delivery efficiency.

• Enterprise-Level Agent Automation Workflows: With the deeply integrated Agent framework, enterprises can build automated workflows for ticket processing, cross-system data synchronization, and intelligent report generation. The model can autonomously plan task steps, call internal APIs, and compile execution results, greatly reducing manual intervention costs in high-repetition business processes such as human resources, financial approvals, and IT operations.

• On-Device Offline Long-Text Processing: In scenarios without network connectivity, mobile devices and in-vehicle systems can directly load on-device models to handle tasks such as book reading summaries, long contract clause reviews, and meeting recording transcription and summarization. The million-token context capacity allows these operations to be performed without splitting the text, ensuring data privacy while improving response speed.

• Intelligent Education and Math Tutoring: Based on its outstanding performance in college entrance exams and AIME math evaluations, the model can support educational applications such as personalized math problem explanations, logical reasoning training, and error analysis. After students submit a problem, the system can generate step-by-step solutions and recommend similar problems, enabling adaptive learning path planning.

• Private Deployment in Domestic Innovation Environments: To meet the strict requirements of government, financial, and energy sectors regarding data not leaving the domain, StarFire X2.5 can be fully privately deployed on domestic chip servers. The model's weights and inference pipeline do not rely on overseas components, complying with security protection and cryptographic evaluation standards, and supporting internal system development such as intelligent customer service, official document writing, and knowledge base Q&A.

• Multilingual International Business: With support for over 200 languages, enterprises can use the model for scenarios such as automatically generating product descriptions for cross-border e-commerce, multilingual customer service chatbots, and overseas market sentiment monitoring. The model directly outputs content in the target language, eliminating the need for a translation intermediary layer and reducing the operational costs of multilingual business.

8. FAQ

Q: Can the on-device model of Spark X2.5 run smoothly on regular smartphones?
A: The 1.7B model can run on smartphones with 4GB or more of memory, with the first token latency approximately in the sub-second range; the 4B model is recommended for use on devices with 8GB of memory. Due to the limitations of on-device computing power, the speed of complex inference tasks will be significantly slower than the cloud API. It is recommended to prioritize on-device solutions for tasks involving long text processing, and to use the cloud API for more challenging inference tasks.

Q: How significant is the capability difference between the cloud model and the on-device model?
A: The cloud 293B/30B MoE models significantly outperform the on-device versions in complex tasks such as mathematical reasoning, code generation, and knowledge-based question answering. While the on-device 4B model performs well in tests like SWE-Bench, there are still noticeable differences in multi-step reasoning consistency, depth of professional knowledge coverage, and stability of instruction-following.

Q: How can I migrate an existing OpenAI API project to Spark X2.5?
A: Spark X2.5 provides an OpenAI-compatible API interface. To migrate, simply replace the base_url with the endpoint provided by the iFLYTEK Open Platform and substitute the api_key with the key obtained from the platform. The request format does not need to be modified. Some advanced parameters (such as logprobs) may have slight differences, so it is recommended to perform full regression testing after migration.

Q: Is the open-source model completely free for commercial use?
A: The on-device models (4B and 1.7B) are licensed under the Apache 2.0 license, which allows free commercial use, modification, and redistribution, but requires the retention of the original copyright notice. The cloud model charges through the API, and is currently in a limited-time free phase. For specific commercial pricing details, please refer to the announcements on the iFLYTEK Open Platform.

Q: Does Spark X2.5 support multimodal input (image, audio)?
A: The current version only supports text input and output and does not natively support image or audio understanding. If you need multimodal capabilities, you can integrate external vision models or speech recognition modules to build a processing pipeline. However, this will increase system complexity and may result in a loss of end-to-end semantic consistency.

9. Project Links

1. GitHub Repository: https://github.com/XHToken/Spark-X2.5
2. Hugging Face Model Library: https://huggingface.co/collections/XHToken/spark-x25 (includes weight files for Spark-X2.5-4B and Spark-X2.5-1.7B)