iFlytek announced the official launch of its upgraded Xunfei Starfire X1, a large-scale deep reasoning model trained with fully domestic computing power. This upgrade represents a comprehensive leap, maintaining leading performance compared to the latest versions of top domestic and international large models such as OpenAI's o3 in overall capabilities, and achieving significant progress in areas such as hallucination control, multilingual ability, and speech simultaneous interpretation, providing users with a smarter, more reliable, and more efficient AI assistant.
The Xunfei Starfire X1 upgrade has significantly improved capabilities in translation, reasoning, text generation, and mathematics, especially in the area of hallucination control. Hallucination is one of the key challenges for the practical application of large models, and Starfire X1 has made notable progress in both factual hallucination control and fidelity hallucination control, significantly enhancing the reliability of large model industry applications. In addition, the multilingual capabilities of Starfire X1 have been greatly expanded, now covering over 130 languages, providing users with barrier-free cross-language communication capabilities.
The speech simultaneous interpretation large model based on the Starfire X1 foundation has achieved significant improvements in translation quality, real-time response, audio perception, and professional depth. The comprehensive quality score of the new version exceeds 90 points, and the first-character response time for Chinese-English simultaneous interpretation has been reduced from 5 seconds in the initial version to as fast as 2 seconds, reaching the industry standard of human-level high-level simultaneous interpretation. In addition, this model has the capability to switch languages without manual intervention, automatically identifying mixed Chinese and English speech and switching translation directions in real time, effectively avoiding communication interruptions caused by language switching operations.
In terms of core technologies, the upgrade of Starfire X1 is attributed to three technological breakthroughs: reinforcement learning technology, data rewriting technology, and hallucination control technology. These technological improvements have effectively reduced the training difficulty of tasks such as mathematical problem-solving, solved the problem of sparse rewards, and alleviated the challenge of obtaining high-quality SFT data. Moreover, they achieve deep alignment with standard answers during the thinking process and response generation of large models, significantly reducing the hallucination rate.
The release of the Xunfei Starfire X1 upgrade not only enhances the overall capabilities of the model but also brings new progress to large models and intelligent agents in industries such as education, healthcare, enterprise applications, code, and research. In the education sector, Starfire X1 supports all scenarios of personalized teaching and learning, significantly improving various capabilities such as homework grading, personalized recommendations, and tutoring. In the medical field, Starfire X1 continues to maintain an industry-leading position in general practice assistance, physical examination report interpretation, and health consultation. In enterprise applications, Starfire X1 significantly improves enterprise knowledge application and industry-depth analysis capabilities through the multimodal knowledge zero-collection system and the Starfire Industry Master product.
In addition, the Xunfei Starfire X1 upgrade offers a better experience for individual users. The new Xunfei Starfire X1 inference model has been launched for the Xunfei Starfire (APP and PC version), optimizing core functions such as text creation, math problem solving, and video production. At the same time, the Xunfei Starfire X1 API has also been launched on the Xunfei Open Platform, providing developers with world-class large model services.