DeepSeek-R1-0528

DeepSeek-R1-0528 is positioned as a May 2025 update to the original DeepSeek R1, refining the same MoE lineage rather than introducing a new family. OpenRouter describes it as a 671B-parameter model with 37B parameters active per inference pass and fully open reasoning tokens, so users can inspect the chain of thought the model produces. Featherless reports a slightly different total parameter figure and emphasizes that the update concentrates on post-training rather than a fresh pre-training run, channeling extra compute into reasoning depth and inference reliability.

The practical gains over the prior R1 are most visible on quantitative reasoning and code tasks, making the model a strong fit for math problem solving, algorithmic coding, agentic workflows, and structured analytical work where step-by-step reasoning matters. Featherless highlights an AIME 2025 jump from 70% to 87.5% accuracy, with the model now spending around 23K tokens per question on average compared to roughly 12K before, alongside MMLU-Redux 93.4, GPQA-Diamond 81.0, and LiveCodeBench 73.3 scores, plus reduced hallucination and improved function calling for tool-using pipelines. OpenRouter also benchmarks the model on par with OpenAI o1, which combined with its open weights gives teams a self-hostable alternative for advanced reasoning workloads.