April 21st.iFLYTEKToday's official announcement.iFlytek Spark X1 New UpgradeIt is claimed to be "the industry's only nationally-produced algorithmic training basedDeep Reasoning Large Models”.
This upgrade has these key messages:
- Achieved significant improvement in the effectiveness of general-purpose tasks such as math, code, logical reasoning, text generation, language comprehension, and knowledge quizzing, with model parameters an order of magnitude smaller than similar models in the industry.Overall results against OpenAI o1 and DeepSeek R1.
- Incorporating more scenarios with complex types of data, the model has made progress in generalization, demonstrated industry-leading capabilities on multiple industry tasks, and further expanded its lead in key industries such as education, healthcare, and justice.
- The first unified model of fast thinking and slow thinking, which supports both modes of thinking by one model at the same timeThe new upgraded model customization and optimization tool chain supports SFT and Reinforcement Learning model customization and optimization solutions with low customization threshold..
- Three core technological innovations -- large-scale multi-stage reinforcement learning training method, unified training method based on fast and slow thinking, and engineering system innovation to guarantee long and stable training based on domestic arithmetic power -- help Starfire X1 to be fully upgraded.
This time, Starfire X1 has been upgraded to breakthrough in several tasks. According to the results of the latest test set, Starfire X1 is fully comparable to OpenAI o1 and DeepSeek R1 in general-purpose tasks, and the results of the latest test set show that Starfire X1 is more effective than OpenAI o1 and DeepSeek R1.Excelled in math, quizzes, etc..
In addition to significant progress in common general-purpose tasks, Starfire X1 has also made progress in model generalization due to the incorporation of more scenarios with complex types of data, and has demonstrated industry-leading capabilities on multiple industry tasks, in theKey industries such as education, healthcare, and justice have further expanded their leading edge.
In order to enable more enterprises to use AI conveniently, the upgraded Starfire X1 provides double support for enterprise AI landing:
- The first fast-thinking and slow-thinking unified model supports two thinking modes at the same time by one model, and the full-blooded version of Starfire X1 can be deployed by 4 cards (Huawei 910B), making private deployment easy.
- New upgraded model customization optimization toolchain thatSupports SFT, reinforcement learning two model custom optimization solutions, 16 cards (Huawei 910B) can complete the industry custom optimization, low customization threshold.
The newly upgraded Cyberfire X1 is now available on the Cyberfire Web or Cyberfire App, allowing users to experience the new effects of deep inference and networked search on general-purpose tasks.
At the same time, Starfire X1 API has also been synchronized on the Xunfei open platform, open services for developers and enterprises.