Recently, the SiliconCloud platform of SiliconBase Flow officially launched the inference acceleration version of DeepSeek-R1-0528 based on domestic computing power. This new version significantly enhances performance, increasing TPM (maximum tokens per minute) to 5 million, meeting high-concurrency demands in complex scenarios. Additionally, the RPM (requests per minute allowed for Pro version R1) has been raised to 30,000, ensuring a smooth user experience.
Users of the original DeepSeek-R1 can automatically enjoy the enhanced model experience without modifying their API parameter configurations. The new version supports features like Function Calling, JSON Mode, Prefix, and FIM. To ensure a smooth transition for enterprise users, the service for the initial version DeepSeek-R1-0120 will continue to be available until June 28th.
DeepSeek-R1-0528 excels in reducing hallucination, decreasing its rate by 45% to 50%. In applications such as rewriting, summarizing, and reading comprehension, the new version provides more accurate and reliable results. For creative writing, this version further optimizes handling of genres like argumentative essays, novels, and prose, capable of producing longer and structurally complete works with a writing style closer to human-like expression.
In terms of tool invocation capabilities, DeepSeek-R1-0528 is comparable to OpenAI o1-high. Moreover, the model demonstrates significant improvements in areas such as front-end code generation and role-playing. Across multiple benchmark tests, the new version performs excellently in mathematics, programming, and general logic, rivaling top international models like o3 and Gemini-2.5-Pro.
User feedback indicates that the new R1 performs more intelligently and humanely. Some developers noted that the version successfully built a word-scoring system during coding challenges, with the generated code and test files running successfully on the first try, making it the second case after o3. Furthermore, DeepSeek-R1-0528 exhibits enhanced language adaptability and reasoning ability, providing users with an enjoyable experience.
Currently, users can call DeepSeek-R1-0528 through the API on the SiliconBase Flow SiliconCloud platform. SiliconBase Flow is committed to providing developers with efficient and stable large model APIs to help users achieve better generative AI applications.
Key Points:
🌟 TPM increased to 5 million, supporting high concurrency needs.
💡 Hallucination rate reduced by 45% to 50%, providing more accurate outputs.
🚀 The new model performs more intelligently with enhanced human-like features.