DeepSeek R1 0528

DeepSeek R1 0528 – Reinforcement Learning-Updated Reasoning Model

DeepSeek R1-0528 is a point release of DeepSeek R1, representing a continuation of DeepSeek's reinforcement learning scaling work on top of the DeepSeek V3 base architecture. It is a dedicated reasoning model — not a hybrid — trained to produce extended chain-of-thought outputs for complex problem-solving.

Architecture

DeepSeek R1-0528 shares the same underlying architecture as DeepSeek R1 and DeepSeek V3: a 671B total parameter, 37B active parameter Mixture-of-Exports dense-attention model. It is initialized from DeepSeek V3-Base.

Reasoning Workflow

R1-0528 operates exclusively in thinking mode: all responses include chain-of-thought reasoning within <think> tags before the final answer. It does not support a non-thinking mode. DeepSeek V3.1 is officially documented as reaching comparable answer quality to R1-0528 while responding more quickly — positioning R1-0528 as the deeper but slower reasoning reference within the DeepSeek family at its release.