Spark

Spark (科大讯飞星火大模型) is a family of large language models developed by iFlyTek (科大讯飞, Kēdà Xùnfēi), a Chinese AI company based in Hefei, Anhui Province. iFlyTek is particularly well-known for its speech recognition and synthesis technology — it is one of the world's leading providers of Chinese speech AI — and the Spark model family extends the company's AI capabilities into large-scale natural language understanding and generation.

Company Background

Founded in 1999 as a spin-off from the University of Science and Technology of China (USTC), iFlyTek has grown to become a major supplier of AI technology to China's government, education system, and enterprise market. The company is publicly listed on the Shenzhen Stock Exchange and has received substantial government contracts for speech-enabled services across public administration, education, healthcare, and judicial systems. iFlyTek's consumer speech recognition technology powers numerous third-party Chinese-language applications.

Spark Model Development

iFlyTek launched its first public Spark LLM model in May 2023, positioning it directly against Baidu's ERNIE Bot in the Chinese market. Subsequent versions — Spark 2.0, 3.0, and 3.5 — progressively improved the model's reasoning, coding, and multilingual capabilities.

Spark 4.0 Turbo emerged as a significant milestone, with iFlyTek reporting that it outperformed GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro across 14 mainstream benchmark tests — a claim reflecting the intense competition among Chinese AI developers to benchmark against Western frontier models.

Spark X1 introduced support for over 130 languages, including comprehensive ASEAN language coverage. The model supports interactive Q&A, problem-solving, content creation, and translation across all supported languages, positioning Spark as a multilingual AI platform with particular relevance to Southeast Asia.

Spark X1.5 (late 2025) was fully trained on domestic Chinese computing platforms, significantly improving training efficiency while extending multilingual processing capabilities. The domestic chip training reflects China's national strategic priority of reducing dependence on US semiconductor supply chains, following NVIDIA export restrictions.

ASEAN Multilingual Model

A particularly notable development is iFlyTek's launch of the Spark ASEAN Multilingual Large Model, a foundation model specifically designed for the ten ASEAN member states and covering ten regional languages: Bahasa Melayu (Malay), Bahasa Indonesia, Thai, Vietnamese, Filipino, Burmese, Khmer, Lao, Malay (Malaysian standard), and Singaporean English. iFlyTek positioned this model as achieving performance comparable to top international standards on ASEAN language benchmarks, addressing a gap left by Western foundation models that are typically under-optimised for Southeast Asian languages.

The ASEAN model reflects iFlyTek's strategy of expanding beyond China by targeting regional markets where its multilingual AI capabilities offer an advantage over English-first Western models.

Product Ecosystem

Beyond the core Spark LLM, iFlyTek offers a range of AI-powered products leveraging its speech-language AI stack. The iFlyTek Recorder is a widely used hardware device combining speech recognition with AI transcription and summarisation. The iFlyTek Learning Machine targets students with AI tutoring capabilities. iFlyTek's medical AI products support clinical documentation and diagnostics in Chinese hospitals. The company's industrial AI platform (Antelope Industrial Large Model) targets manufacturing and process industries.

See Also

References

References

1. iFlyTek. (2025). iFlyTek Spark LLM. global.xfyun.cn/products/SparkLLM.
2. Pandaily. (2025). iFlytek unveils Spark X1.5 deep reasoning model. pandaily.com.
3. iFlyTek. (2025). SPARK multi-language large model released. iflytek.com/en.
4. iFlyTek. (2025). iFlyTek showcases comprehensive AI innovations at WAIC 2025. iflytek.com/en.
5. WIC Internet. (2025). iFlytek: Using AI to build a better world. wicinternet.org.