DeepSeek R1 (May 2025 update) is the refined reasoning powerhouse — the same open-weight, MIT-licensed model that challenged o1-preview, now with improved consistency and instruction following.
DeepSeek-R1 (May 28) is an AI model released May 28, 2025 by DeepSeek, a leading organization in large language model research. May 2025 update to DeepSeek-R1 — improved reasoning depth, better code generation, and enhanced benchmark scores. With 671B total (37B active via MoE) and a 128K tokens context window, DeepSeek-R1 (May 28) is designed to excel across a range of demanding tasks — from software engineering and mathematical problem-solving to multilingual content generation and long-document analysis.
DeepSeek-R1 (May 28) employs a Mixture-of-Experts (MoE) architecture with 671B total (37B active via MoE). In a MoE design, only a subset of "expert" sub-networks activates for each token during inference — DeepSeek-R1 (May 28) has 671B total (37B active via MoE) total parameters but uses only a fraction per token. This allows the model to store significantly more learned knowledge than a comparably-priced dense model. MoE architectures have proven effective at reaching frontier-level capability while keeping serving costs manageable — a key reason DeepSeek-R1 (May 28) can be offered at the scale that platforms like LITHOVEX AI make possible.
As of 2026-10-16, DeepSeek-R1 (May 28) by DeepSeek demonstrates measurable performance on standard AI evaluation benchmarks. On AIME 2025 — which tests advanced competition mathematics at the highest US high school level — it achieves 87.5%. For software engineering tasks, DeepSeek-R1 (May 28) scores 57.6% on SWE-bench Verified, which tests ability to resolve real GitHub issues in production Python repositories. In mathematical reasoning, DeepSeek-R1 (May 28) records 87.5% on AIME 2025. Across all 4 benchmarks tracked, the model shows consistent capability in its primary task domains. These scores place DeepSeek-R1 (May 28) — at 671B total (37B active via MoE) of parameters — in a competitive position within its model class. Note that benchmark scores reflect controlled evaluation conditions; real-world performance varies by task, prompt design, and domain specificity.
The primary strengths of DeepSeek-R1 (May 28) are: Significant improvement over original R1 — SWE-bench 49.2% → 57.6%; 87.5% AIME 2025 — near-human on elite US mathematics competition; Reduced overthinking — more concise reasoning chains; Same MIT license as original. These make it particularly well-suited for Software engineering tasks requiring real-world code changes, Competition mathematics at AIME/Olympiad level, Drop-in improvement wherever R1 was previously used. Users working on complex, multi-step problems benefit most from DeepSeek-R1 (May 28)'s ability to maintain coherence across long inputs and produce structured, verifiable outputs.
Like all current AI systems, DeepSeek-R1 (May 28) has limitations: Same massive infrastructure requirements as R1; Knowledge cutoff May 2025. For production deployments, validating outputs against authoritative sources — especially for domain-specific or time-sensitive information — remains best practice. DeepSeek continues to improve DeepSeek-R1 (May 28) through ongoing research.
DeepSeek-R1 (May 28) is available on LITHOVEX AI with a free account. Unlike direct API access — which requires billing setup, API key management, and per-token charges — LITHOVEX AI provides a seamless browser interface with no setup required, making DeepSeek-R1 (May 28) accessible to students, researchers, developers, and early-stage teams globally.
As of 2026-10-16, DeepSeek-R1 (May 28) by DeepSeek demonstrates measurable performance on standard AI evaluation benchmarks. On AIME 2025 — which tests advanced competition mathematics at the highest US high school level — it achieves 87.5%. For software engineering tasks, DeepSeek-R1 (May 28) scores 57.6% on SWE-bench Verified, which tests ability to resolve real GitHub issues in production Python repositories. In mathematical reasoning, DeepSeek-R1 (May 28) records 87.5% on AIME 2025. Across all 4 benchmarks tracked, the model shows consistent capability in its primary task domains. These scores place DeepSeek-R1 (May 28) — at 671B total (37B active via MoE) of parameters — in a competitive position within its model class. Note that benchmark scores reflect controlled evaluation conditions; real-world performance varies by task, prompt design, and domain specificity.
Mixture-of-Experts reasoning model: 671B total / 37B active, improved May 2025 checkpoint
The training pipeline for DeepSeek-R1 (May 28) involves multiple stages: large-scale pretraining on diverse multilingual text and code corpora to develop broad world knowledge, followed by supervised fine-tuning on instruction-response pairs, and alignment techniques such as RLHF or DPO. Each stage refines the model's helpfulness, accuracy, and safety. DeepSeek's research team — based in China — has optimized DeepSeek-R1 (May 28) for the specific capability profile described above.
Inference efficiency: DeepSeek-R1 (May 28) at 671B total (37B active via MoE) is designed to deliver quality outputs at practical serving speeds. Its 128K tokens context window enables processing entire documents, multi-file codebases, and extended conversation histories — particularly valuable for contract review, technical documentation analysis, and codebase-level AI assistance.
Getting started with DeepSeek-R1 (May 28) on LITHOVEX AI takes under two minutes. Follow these steps:
For best results with DeepSeek-R1 (May 28): provide specific, context-rich prompts; include examples when generating code; specify output format when structure matters; break complex tasks into focused sub-questions. Since DeepSeek-R1 (May 28) supports extended reasoning, prompting it to "think step by step" or "show your reasoning" activates deeper chain-of-thought analysis for math, logic, and complex problems.
Complex mathematical reasoning and theorem proving
Scientific research analysis requiring chain-of-thought
Legal and regulatory document analysis with step-by-step logic
Financial modeling and quantitative strategy development
Academic research that benefits from visible reasoning chains
DeepSeek is a Chinese AI research company founded in 2023, backed by High-Flyer — one of China's leading quantitative hedge funds. Despite its brief history, DeepSeek has made an outsized global impact with open-weight releases, particularly the DeepSeek-R1 and DeepSeek-V3 families. The company's research philosophy prioritizes efficiency and transparency: models are released under MIT licenses, and detailed technical reports accompany each release, influencing the broader AI research community. DeepSeek's engineering team demonstrated that frontier-level performance could be achieved at dramatically lower training costs than Western competitors — a revelation that reshaped industry assumptions about AI development economics. As of 2025, DeepSeek's models rank among the most-used open AI models globally.
Funding: Backed by High-Flyer Capital Management
Founded: 2023, headquartered in China.
DeepSeek-R1 (May 28) by DeepSeek competes in a field with several strong alternatives. Below is a detailed side-by-side comparison against commonly listed alternatives, using benchmark data and capability profiles in the catalog.
Both DeepSeek-R1 (May 28) and the models above are available free on LITHOVEX AI — test them side by side on your real tasks before committing to a subscription with any provider.
DeepSeek-R1 (May 28) is released under MIT, which means you can download and run it on your own hardware. Below are the quickest ways to get started.
LITHOVEX AI runs DeepSeek-R1 (May 28) on cloud infrastructure, so you can use it instantly in your browser without downloading anything, configuring drivers, or managing VRAM. Create a free account and start in under 60 seconds.
DeepSeek R1 0528 continued the momentum of the original R1 release, maintaining its position as the global standard for open-weight reasoning AI. Used by research institutions and enterprises across 50+ countries who need o1-class reasoning without proprietary lock-in.
DeepSeek-R1 (May 28) by DeepSeek has the following version history and notable updates:
For the complete changelog and latest release notes, visit the official DeepSeek model page.