DeepSeek R1-0528

2025-05-28 · Publisher : DeepSeek · Domain : Artificial Intelligence

DeepSeek releases R1-0528, an upgrade to R1 that doubles the reasoning depth and raises AIME 2025 from 70.0 to 87.5, with a version distilled into Qwen3 8B.

### Context Four months after R1, the closed-model market had moved forward with OpenAI o3 and Gemini 2.5 Pro. DeepSeek responded with an update to R1 rather than a new name: same architecture, but more post-training compute and new optimization mechanisms. The goal is to keep an open model on par with the best closed reasoning models. ### What the model brings - Deeper thinking: about 23K tokens per question on AIME versus 12K for the previous version (model card). - Clear gains: GPQA Diamond from 71.5 to 81.0, LiveCodeBench from 63.5 to 73.3, SWE Verified from 49.2 to 57.6, HMMT 2025 from 41.7 to 79.4. - A reduced hallucination rate, function calling and JSON output supported, and a better experience for "vibe coding". - The system prompt is now accepted, and forcing the thinking tag is no longer necessary. - A distilled model, DeepSeek-R1-0528-Qwen3-8B, announced as the best among open models on AIME 2024. ### Architecture and specifications - Type: Mixture-of-Experts, 685 billion parameters in total (Hugging Face value). Active parameters and number of experts: not included in the model card consulted, which states the R1 architecture. - Context: not disclosed; the maximum generation length used for evaluations is 64K tokens. - Modalities: text. Data and knowledge cutoff: not disclosed. - License: MIT, commercial use and distillation expressly permitted. - Weights: huggingface.co/deepseek-ai/DeepSeek-R1-0528, and DeepSeek-R1-0528-Qwen3-8B for the distilled version. ### Cost and access | Item | Detail | |---|---| | API | Identifier `deepseek-reasoner`, "no change in usage" | | Pricing at launch | Not disclosed in the post (R1 pricing was 0.14 / 0.55 / 2.19 USD per million tokens) | | Self-hosting hardware | Not disclosed | ### Benchmarks | Benchmark | R1 (January) | R1-0528 | |---|---|---| | MMLU-Pro (EM) | 84.0 | 85.0 | | GPQA Diamond (Pass@1) | 71.5 | 81.0 | | Humanity's Last Exam (Pass@1) | 8.5 | 17.7 | | LiveCodeBench (Pass@1) | 63.5 | 73.3 | | SWE Verified | 49.2 | 57.6 | | AIME 2024 / 2025 | 79.8 / 70.0 | 91.4 / 87.5 | | HMMT 2025 | 41.7 | 79.4 | For the distilled variant (Qwen3-8B), the same card compares it with closed models: | Model | AIME 24 | AIME 25 | HMMT Feb 25 | GPQA Diamond | LiveCodeBench | |---|---|---|---|---|---| | R1-0528-Qwen3-8B | 86.0 | 76.3 | 61.5 | 61.1 | 60.5 | | Gemini-2.5-Flash-Thinking-0520 | 82.3 | 72.0 | 64.2 | 82.8 | 62.3 | | o3-mini (medium) | 79.6 | 76.7 | 53.3 | 76.8 | 65.9 | Source: DeepSeek-R1-0528 model card on Hugging Face, scores published by the vendor. ### Versus closed models The vendor writes that R1-0528's performance "approaches" that of o3 and Gemini 2.5 Pro, although the text table of the card does not give the scores of these two models (the comparison is in an image chart). The verifiable figures concern the 8-billion-parameter distilled variant: it surpasses o3-mini (medium) on AIME 2024 (86.0 versus 79.6) and Gemini 2.5 Flash Thinking on AIME 2024 (86.0 versus 82.3). It stays behind on GPQA Diamond (61.1 versus 76.8 for o3-mini and 82.8 for Gemini Flash). The full model gains nearly ten points on GPQA in four months. This depth is paid for in output tokens: 23K per question on AIME versus 12K previously. ### Key takeaways R1-0528 proves that a post-training update can move an open model from one level to the next without changing the architecture. The Qwen3 8B version carries part of this reasoning over to a small model.

Tags : LLM, Open Source, DeepSeek, MIT, MoE, Raisonnement, Distillation