If you need an AI model that can reason through multi-step problems rather than just pattern-matching, DeepSeek R1 is the open-source answer. It scores 97.3% on MATH-500 in DeepSeek’s own paper. DeepSeek says its May 2025 update (R1-0528) approaches the performance of models like O3 and Gemini 2.5 Pro. The best part is that its distilled variants run on hardware you already own.
We have been testing R1 at VORLUX AI for code review, financial analysis, and compliance document reasoning. Here is what we found.
What Makes R1 Different: Chain-of-Thought Reasoning
Most language models give you an answer. DeepSeek R1 shows you its thinking. When you ask it a complex question, it produces an explicit chain-of-thought before arriving at its conclusion, visible in the output as a “thinking” block.
This matters for business use cases because:
- Auditability: You can verify how the model reached its conclusion, not just what it concluded. For legal analysis or financial modeling, this is the difference between a useful tool and a black box.
- Error detection: When the reasoning is visible, mistakes become obvious. A wrong step in the chain stands out, while a wrong final answer from a standard model gives you nothing to debug.
- Trust building: Showing clients the model’s reasoning process builds confidence in local AI deployments. It’s no longer “the AI said so”, it’s “here’s the analysis.”
Benchmarks: Where R1 Stands
DeepSeek R1’s full 671B Mixture-of-Experts model delivers results that would have been unthinkable for open-source two years ago. All figures below are DeepSeek’s own evaluations, from the R1 model card and paper (pass@1), plus the R1-0528 model card for AIME 2025:
| Benchmark | DeepSeek R1 (Full) | R1 32B Distill | R1 14B Distill | GPT-4o (0513) |
|---|---|---|---|---|
| MATH-500 | 97.3% | 94.3% | 93.9% | 74.6% |
| AIME 2024 | 79.8% | 72.6% | 69.7% | 9.3% |
| AIME 2025 | 70.0% (R1-0528: 87.5%) | — | — | — |
xychart-beta
title "DeepSeek R1 vs Competitors — MATH-500 Score"
x-axis ["R1 Full (671B)", "GPT-4o (0513)", "R1 32B Distill", "R1 14B Distill"]
y-axis "Score (%)" 50 --> 100
bar [97.3, 74.6, 94.3, 93.9]The key insight: on DeepSeek’s own evaluation, the 14B distilled variant scores 93.9% on MATH-500, well ahead of the 74.6% it reports for GPT-4o, and adds a visible chain of thought. Treat vendor benchmarks as a shortlist, not a verdict: test on your own documents before you commit.
How to Run R1 Locally with Ollama
Getting started takes one command:
# 14B — fits on Mac Mini M4 (16GB)
ollama pull deepseek-r1:14b
# 32B — needs 32GB+ unified memory
ollama pull deepseek-r1:32b
# Run with a reasoning prompt
ollama run deepseek-r1:14b "A company has EUR 50,000 to invest in AI infrastructure. Compare the 3-year TCO of cloud API usage at EUR 800/month versus a one-time local deployment with ongoing maintenance. Include opportunity cost of the upfront investment at 5% annual return."The model will output its thinking process first, then the final answer. This is normal: it’s the chain-of-thought at work.
Hardware Requirements
| Variant | Parameters | Download size (Ollama) | Best For |
|---|---|---|---|
| R1 1.5B | 1.5B | 1.1 GB | Quick classification, simple Q&A |
| R1 7B | 7B | 4.7 GB | General reasoning, drafting |
| R1 14B | 14B | 9.0 GB | Sweet spot for SME deployment |
| R1 32B | 32B | 20 GB | Complex analysis, code review |
| R1 Full | 671B MoE | 404 GB | Research, maximum quality; multi-GPU only |
Speed depends on your memory bandwidth and on how long the model thinks; measure it on your own machine with ollama run deepseek-r1:14b --verbose.
For most business deployments, the 14B distill is the sweet spot. It fits on a Mac Mini M4 with 16GB and delivers strong reasoning at interactive speeds. If your hardware has 32GB+ memory, the 32B variant offers notably better quality.
Real Use Cases at VORLUX AI
We run DeepSeek R1 14B for tasks that require genuine reasoning, not just text generation:
Contract analysis: Feed it a 20-page service agreement and ask “What are the three most one-sided clauses in this contract and why?” The chain-of-thought output walks through each clause, compares terms to standard practice, and flags specific risks. In our use, the analysis goes deeper than what we got from Gemma 2 on the same prompt.
Financial modeling: “Given these 12-month revenue projections, what’s the break-even point if we add a EUR 2,400/month developer salary in month 4?” R1 doesn’t just calculate: it identifies assumptions, checks edge cases, and warns about scenarios you didn’t ask about.
Code debugging: When our n8n code review workflow encounters a complex bug, R1’s chain-of-thought traces through the execution path step by step, identifying the exact point where logic diverges from intent.
R1 vs DeepSeek V3: When to Use Which
We run both DeepSeek models. Here’s how we decide:
| Task Type | Best Model | Why |
|---|---|---|
| Multi-step reasoning | R1 | Chain-of-thought is essential |
| Fast text generation | V3 | Higher throughput, no thinking overhead |
| Code review | R1 | Traces logic paths, catches subtle bugs |
| Content drafting | V3 | Speed matters more than deep reasoning |
| Compliance analysis | R1 | Auditable reasoning chain |
| Customer Q&A | V3 | Quick responses, no thinking delay |
For a deeper look at DeepSeek V3, see our DeepSeek V3 review.
The Privacy Advantage
Every chain-of-thought step happens on your hardware. When R1 reasons through a financial model or analyzes a legal contract, that reasoning, including any sensitive data it references, never leaves your building.
This is particularly relevant under GDPR and the EU AI Act, in general application since 2 August 2026 (Regulation (EU) 2024/1689). Automated decision-making on personal data requires transparency about how decisions are reached. R1’s visible reasoning chain is the technical implementation of that transparency requirement.
Compare this to sending the same contract to a cloud API: the data leaves your premises, gets processed on servers you don’t control, and the reasoning is a black box. With R1 running locally, the entire process is auditable, contained, and yours.
The Bottom Line
DeepSeek R1 closes the reasoning gap between open-source and proprietary models. The 14B distilled variant delivers chain-of-thought reasoning that beats GPT-4o on DeepSeek’s published math benchmarks: running on a EUR 700 Mac Mini with zero per-query costs.
For European SMEs dealing with contracts, compliance, financial analysis, or code (tasks where how the AI thinks matters as much as what it says) R1 is the model to deploy.
Ready to deploy DeepSeek R1 in your business? Schedule a free 15-minute assessment to see how chain-of-thought reasoning can transform your workflows.
More model reviews: Best Local LLM Models Q2 2026 | DeepSeek V3 Review | Phi-4 Review
Sources: DeepSeek R1 on Ollama | DeepSeek R1 model card | DeepSeek-R1 paper (arXiv 2501.12948) | DeepSeek R1-0528 model card
Related reading
- Best Local LLM Models for Q2 2026: Practical Comparison for SMEs
- Cloud vs Local AI: Real Cost Analysis for Spanish SMEs in 2026
- Local AI Readiness Checklist: Is Your Business Ready to Run AI On-Premise?
Work with us
We size the model and the machine by measuring, not by guessing. If you want to see your own task running on real hardware, book a 15-minute call or see how we work in consulting.