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

What is iFlytek Spark X2

iFlytek Spark X2 is iFlytek's new-generation large model trained on fully domestic computing power. It adopts a 293B MoE sparse architecture, with general capabilities benchmarked against the world's top level, and capabilities such as mathematical reasoning and 130+ multilingual support comparable to the world's best. Through engineering innovation, the model achieves deployment on a single Ascend server, with inference performanceX1.5improved by 50%. It is the industry's first to feature personalized learning capabilities driven by error causes, deeply empowering industries such as education, healthcare, and automotive, with vague intent understanding jumping from unusable to basically usable. The model's latest capabilities canbe experienced inthe iFlytek Spark web version and APP, and the new API has also been launched on the iFlytek Open Platform。

Main features of iFlytek Spark X2

• Mathematical reasoning: The model can solve highly difficult competition problems and provide detailed solution processes.
• Multilingual processing: It supports more than 130 languages and maintains industry-leading performance in key languages such as those of Latin America and ASEAN.
• Code generation: It has excellent code understanding and generation capabilities and can assist programming development.
• Long-text understanding: The model can handle complex long-context tasks.
• Agent building: It supports complex task planning that "can think and execute" and multi-tool combination invocation.
• Education empowerment: It achieves step-level correction, precise error cause localization, and heuristic explanations, and is the industry's first to feature personalized learning capabilities driven by error causes.
• Healthcare empowerment: It provides intelligent health analysis, report interpretation, medication review, and multi-round proactive consultation services.
• Automotive interaction: It achieves vague intent understanding and high-EQ intelligent cockpit human-machine interaction.
• Enterprise applications: It provides bidding and procurement compliance detection, supplier verification, and in-depth industry analysis functions.
• Personal assistant: It covers scientific research assistance and full-process job-seeking services, including resume optimization and mock interviews.

Technical principles of iFlytek Spark X2

• MoE sparse architecture: Spark X2 adopts a 293B-parameter mixture-of-experts model architecture, improving parameter usage efficiency through a sparse activation mechanism, reducing computational overhead while ensuring model capabilities.
• Model compression technology: It uses weight quantization and low-precision KV Cache technology to achieve lightweight model compression, significantly reducing memory usage and improving inference speed, enabling the large model to run efficiently on domestic computing power platforms.
• Parallel deployment solution: It adopts VTP virtual tensor parallelism and a hierarchical communication mechanism to complete large EP parallel deployment on domestic computing power platforms, enabling the large model to run on a single Ascend server.
• Training-inference calibration algorithm: Aiming at the inconsistency between training and inference probability distributions in reinforcement learning for MoE models, an adaptive calibration algorithm was developed, greatly improving the accuracy and stability of reinforcement learning training.
• Data synthesis method: A recursive high-difficulty data synthesis scheme was designed, gradually correcting reasoning errors through multiple rounds of iterative derivation, continuously building scarce high-quality data, and improving the model's deep reasoning accuracy.
• High-throughput sampling optimization: A multi-stage sampling scheme with P/D two-stage separation was adopted to solve the efficiency interference problem under high-throughput sampling in reinforcement learning on domestic platforms, improving training efficiency by 10%.

How to use iFlytek Spark X2

• Web version experience: Visit theiFlytek Sparkofficial website to use it online without downloading, supporting text chat, voice interaction, document upload, and other features.
• Mobile use: DownloadiFlytek Spark APP version 5.2.0, which supports one-tap voice chat and lets you call the AI assistant anytime, anywhere, with especially enhanced research assistant and job-seeking assistant features.
• API integration development: Developers can use the iFlytek Open Platform https://xinghuo.xfyun.cn/sparkapi to access the Spark X2 API, supporting integration of the model's capabilities into their own applications.

Application scenarios of iFlytek Spark X2

• Education and learning: Applied to AI learning devices to achieve step-by-step homework correction, precise error cause identification, and heuristic explanations, building a personalized learning loop and improving students' interest and efficiency in learning.
• Smart cockpit: Empowers automotive voice interaction systems to achieve fuzzy intent understanding and high-EQ responses, enabling cars to accurately understand users' "subtext" like an old friend, with industry-leading real-vehicle evaluations.
• Enterprise agents: Applied to scenarios such as bidding and procurement compliance detection, supplier verification, and bid anomaly monitoring, improving efficiency by more than 3 times and compressing development cycles from several days to the minute level.
• Personal efficiency improvement: The research assistant supports the entire paper-writing process, while the job-seeking assistant automatically gathers recruitment information, optimizes resumes, and simulates interviews, meeting young people's needs for learning and career development.