iFlytek Spark X2 - iFlytek's new generation large-scale model

What is Xunfei Xinghuo X2?

iFlytek Spark X2 is a new generation of large-scale models trained using domestically produced computing power. It employs a 293B MoE sparse architecture, and its general capabilities are on par with top international standards. Its mathematical reasoning and 130+ multi-language capabilities are comparable to the best international standards. Through engineering innovation, the model is deployed on a single Ascend server, achieving a 50% improvement in inference performance compared to X1.5. It features an industry-first personalized learning capability that delves into error causes, deeply empowering industries such as education, healthcare, and automotive, and elevating fuzzy intent understanding from unusable to essentially usable. The model's latest capabilities include...existThe iFlytek Spark web version and app experience, along with the new API, are now available on the iFlytek Open Platform..

Main functions of iFlytek Starfire X2

• Mathematical reasoningThe model can solve highly difficult competition problems and provides detailed solutions.
• Multilingual processingIt supports more than 130 languages and maintains an industry-leading position in key languages such as Latin American and ASEAN languages.
• Code generationIt possesses excellent code comprehension and generation capabilities, which can assist in programming development.
• Long text comprehensionThe model can handle complex tasks with long contexts.
• Intelligent agent constructionIt supports complex task planning and multi-tool combination invocation that can "think and execute".
• Education empowermentIt enables step-by-step correction, precise error identification, and heuristic explanation, and is the first in the industry to provide personalized learning capabilities that integrate error analysis.
• Healthcare EmpowermentIt provides intelligent health analysis, report interpretation, medication review, and multiple rounds of proactive consultation services.
• Car InteractionTo achieve fuzzy intent understanding and highly emotionally intelligent cockpit human-computer interaction.
• Enterprise ApplicationsIt provides functions for procurement compliance testing, supplier verification, and in-depth industry analysis.
• Personal AssistantIt covers the entire process of scientific research assistance and job search services, including resume optimization and mock interviews.

Technical Principles of iFlytek Starfire X2

• MoE Sparse ArchitectureThe Spark X2 adopts a hybrid expert model architecture with 293B parameters. It improves the efficiency of parameter utilization through a sparse activation mechanism, reducing computational overhead while ensuring model capabilities.
• Model compression technologyThe model is lightweight and compressed by using weight quantization and low-precision KV Cache technology, which significantly reduces memory usage and improves inference speed, enabling large models to run efficiently on domestic computing platforms.
• Parallel deployment schemeBy adopting VTP virtual tensor parallelism and hierarchical communication mechanism, large EP parallel deployment is completed on the domestic computing power platform, enabling a large model to run on a single Ascend server.
• Training and calibration algorithmTo address the inconsistency between training and inference probability distributions in MoE model reinforcement learning, an adaptive calibration algorithm was developed, which significantly improves the accuracy and stability of reinforcement learning training.
• Data synthesis methodsThe design of a recursive high-difficulty data synthesis scheme gradually corrects reasoning errors through multiple rounds of iterative derivation, continuously builds scarce high-quality data, and improves the accuracy of deep reasoning in the model.
• High-throughput sampling optimizationThe multi-stage sampling scheme with P/D two-stage separation is adopted to solve the efficiency interference problem in the case of high-throughput sampling of reinforcement learning on domestic platforms, and the training efficiency is improved by 10%.

How to use iFlytek Starfire X2

• Web version experienceVisit the iFlytek Spark website to use it online without downloading. It supports functions such as text conversation, voice interaction, and document upload.
• Mobile useDownload the iFlytek Spark APP version 5.2.0, which supports one-click voice dialogue and allows you to call up the AI assistant anytime, anywhere. The research assistant and job search assistant functions have been specially enhanced.
• API Integration DevelopmentDevelopers can access the Spark X2 API through the iFlytek Open Platform (https://xinghuo.xfyun.cn/sparkapi) to integrate model capabilities into their own applications.

Application scenarios of iFlytek Spark X2

• Education and LearningIt is applied to AI learning machines to achieve step-by-step homework correction, accurate error identification, and heuristic explanation, thereby building a personalized learning loop and improving students' learning interest and efficiency.
• Smart CockpitEmpowering automotive voice interaction systems, enabling fuzzy intent understanding and high-EQ responses, allowing cars to accurately understand users' "subtext" like an old friend, leading the industry in real-vehicle evaluations.
• Enterprise Intelligent AgentIt can be applied to scenarios such as procurement compliance testing, supplier verification, and bid anomaly monitoring, improving efficiency by more than 3 times and reducing the development cycle from several days to minutes.
• Personal efficiency improvementResearch assistants provide full-process support for paper writing, while job search assistants automatically collect recruitment information, optimize resumes, and conduct mock interviews, meeting the learning and career development needs of young people.