DeepSeek R1 0528 is a May 2025 refresh of the original DeepSeek R1 reasoning model, positioned as an open-weights alternative to leading closed systems. The release notes highlight improved benchmark performance, stronger front-end generation, fewer hallucinations, and newly added JSON output and function calling, while keeping the same API contract. The model is delivered with openly published weights and is widely routed through third-party providers, making it attractive for teams that want frontier-style reasoning without lock-in to a single vendor.

Under the hood, R1 0528 uses a Mixture-of-Experts design with 671B total parameters and 37B active per inference pass, which is why it can be efficient at serving time despite its scale. It supports a long context window of roughly 164K tokens, handles text input and output, and exposes thinking tokens as part of its open reasoning surface. Together, those traits make it a practical fit for complex analytical tasks, agentic pipelines that need structured outputs or tool use, and research or product teams that want to self-host a strong reasoning model.