DeepSeek updates R1 with JSON output and function calling

DeepSeek-R1-0528 is an upgraded release of DeepSeek's open reasoning model. If you already use R1 for logic-heavy tasks, this update brings the missing pieces developers have been waiting for: native tool use and structured outputs.

What it can do

The biggest addition is native support for structured JSON output and function calling. Developers no longer need to write custom parsers or hack prompts just to extract clean data from a reasoning trace. You can now wire the model directly into multi-step agent pipelines and external tool environments.

DeepSeek also reports higher benchmark scores across the board, better capability on front-end tasks, and a noticeable drop in hallucinations.

You can try the new model right away in the web chat. For API developers, DeepSeek kept the integration seamless with zero breaking changes to existing endpoints. If you want to host it yourself, the model weights are already live on Hugging Face.

Why it matters

Reasoning models are great at solving hard problems, but without reliable JSON formatting and function calling, they are clumsy to use inside production software. DeepSeek-R1-0528 fixes that exact bottleneck.

In our view, dropping the open weights immediately while keeping API backwards compatibility makes this an easy upgrade. If you build agents or rely on structured reasoning pipelines, it is well worth testing right away.