DeepSeek R1-0528

Rolled out May 28, 2025, DeepSeek-R1-0528 is a minor-version upgrade of the open-weights reasoning model that shook markets in January 2025. Using more post-training compute and algorithmic refinements, it deepened chain-of-thought reasoning (average ~23K reasoning tokens per question versus ~12K), lifting AIME 2025 accuracy from 70% to 87.5% and LiveCodeBench from 63.5% to 73.3%, with reduced hallucinations and new JSON-output plus function-calling support. Like the original, the full 671B-parameter MoE weights are MIT-licensed on Hugging Face — free for commercial use and distillation. A distilled R1-0528-Qwen3-8B variant runs on a single GPU.\n\nLimitations: the full model needs roughly a dozen 80GB GPUs to run, so most users rely on the API or distilled versions. DeepSeek's own release note warns that complex reasoning tasks consume substantially more tokens than the original R1. It remains a reasoning-first model — weaker at casual chat polish and tool-heavy agentic workflows than purpose-built agent models.

DeepSeek-reported: 87.5% AIME 2025, 81.0% GPQA, 73.3% LiveCodeBench