DeepSeek R1-0528

DeepSeek-R1-0528 is a point release in the DeepSeek-R1 family, developed by DeepSeek-AI as a foundation model focused on automated reasoning, coding, and general language understanding. It builds on the original DeepSeek-R1 line, continuing the lineage documented in the accompanying research paper (arXiv:2501.12948) and the open-source code released at the DeepSeek-AI GitHub repository. The release post frames this build as a refinement of the reasoning-first design, emphasizing improved benchmark performance, enhanced front-end capabilities, and a meaningful reduction in hallucinations compared with earlier versions.

Beyond its reasoning roots, DeepSeek-R1-0528 introduces practical developer-facing features, notably native support for structured JSON output and function calling, while keeping the same thinking-mode API usage patterns as before. Independent benchmark comparisons shared with the release place it competitively against other advanced models on academic and code evaluations such as AIME, GPQA Diamond, and LiveCodeBench. The model is distributed as open weights under the MIT License via the deepseek-ai/DeepSeek-R1-0528 repository on Hugging Face, making it well suited for self-hosted deployments, research into large-scale reasoning, and applications that need both chain-of-thought style answers and reliable structured outputs.