DeepSeek R1 Local Setup Ollama: Complete Installation & Optimization Guide

Quick Answer: To run DeepSeek R1 locally with Ollama, run ollama run deepseek-r1:14b for 16GB GPUs or ollama run deepseek-r1:8b for 8GB GPUs. For maximum reasoning fidelity, configure a custom Modelfile setting temperature 0.6 and top_p 0.95, preserving the <think> reasoning traces without artificial system prompt overrides.

Last Updated: September 3, 2026 | Reviewed by Senior Systems Architect

Key Takeaways

• Model Variants: DeepSeek R1 distilled checkpoints are available in 1.5B, 7B, 8B, 14B, 32B, and 70B sizes, with the 14B Q4_K_M model offering the best performance-to-VRAM balance on consumer hardware.
• VRAM Thresholds: The 8B model requires 5.8 GB VRAM; the 14B model requires 9.6 GB VRAM; the 32B model requires 20.2 GB VRAM. Sizing can be verified in our VRAM Requirements Calculator.
• Prompting Constraint: Do not inject aggressive system instructions instructing the model to suppress thinking; doing so disrupts chain-of-thought mathematical and coding derivation.
• Production Routing: Integrate Ollama with agent orchestrators via our Custom MCP Server Tutorial or scale to multi-user clusters with vLLM Serving Benchmarks.

---

1. DeepSeek R1 Local Hardware Requirements Table

Model Size	Quantization	Required VRAM	Minimum Recommended GPU	Tokens / Sec (RTX 4090)
DeepSeek-R1-1.5B	Q4_K_M	1.8 GB	Any Modern iGPU / GTX 1650	145 t/s
DeepSeek-R1-7B	Q4_K_M	5.2 GB	RTX 3060 (12GB) / RTX 4060	88 t/s
DeepSeek-R1-8B	Q4_K_M	5.8 GB	RTX 3060 (12GB) / Apple M2 (16GB)	82 t/s
DeepSeek-R1-14B	Q4_K_M	9.6 GB	RTX 4070 (12GB) / RTX 3080	58 t/s
DeepSeek-R1-32B	Q4_K_M	20.2 GB	RTX 3090 (24GB) / RTX 4090	36 t/s
DeepSeek-R1-70B	Q4_K_M	43.5 GB	2x RTX 3090 (48GB) / Mac Studio 64GB	19 t/s

---

2. Step-by-Step Installation Commands

Ensure Ollama is updated to the latest binary release supporting QwQ and DeepSeek architecture optimizations:

# 1. Pull and run the balanced 14B reasoning model
ollama run deepseek-r1:14b

# 2. Test chain-of-thought reasoning in terminal
>>> "Solve the following problem step-by-step: Write a Python function to find the longest palindromic substring."

Ollama automatically initializes GPU layer offloading according to the Ollama Official Release Notes.

---

3. Creating an Optimized Custom Modelfile

Standard default configurations lack parameter tuning for mathematical reasoning. Create a dedicated Modelfile:

FROM deepseek-r1:14b

# Set optimal sampling parameters for reasoning tasks
PARAMETER temperature 0.6
PARAMETER top_p 0.95
PARAMETER top_k 40
PARAMETER num_ctx 32768

# Preserve native reasoning token templates
TEMPLATE """{{ if .System }}<｜System｜>{{ .System }}{{ end }}{{ range .Messages }}{{ if eq .Role "user" }}<｜User｜>{{ .Content }}<｜Assistant｜>{{ else if eq .Role "assistant" }}<｜thought｜>{{ .Content }}{{ end }}{{ end }}"""

Compile and register the model:

ollama create r1-coder -f ./Modelfile
ollama run r1-coder

---

4. Benchmarking Accuracy and Speed

According to technical disclosures in the DeepSeek R1 Research Paper and the DeepSeek GitHub Architecture Repository, test-time compute scaling delivers parity with proprietary models across AIME and MATH-500 benchmarks.

For developers deploying on Apple hardware, review our unified memory benchmarks in the Mac Studio M4 Max Review.

---

5. Frequently Asked Questions (FAQ)

Why does DeepSeek R1 output <think> blocks?

The <think> tags contain the raw chain-of-thought reflection where the model verifies assumptions, explores edge cases, and self-corrects before providing the final answer. Removing or masking these tokens degrades output accuracy on logic benchmarks.

How do I expose DeepSeek R1 as an OpenAI-compatible API?

Ollama automatically serves an OpenAI-compatible endpoint on port 11434. Point your applications to http://localhost:11434/v1 using deepseek-r1:14b as the model name.

Can I run the full 671B DeepSeek R1 model locally?

Running the un-distilled 671B MoE model requires approximately 380 GB of VRAM even at 4-bit quantization, necessitating an 8x H100 datacenter cluster or multiple Mac Studio 192GB nodes linked via high-speed cluster networking.

FROM deepseek-r1:14b
PARAMETER temperature 0.6
PARAMETER top_p 0.95
PARAMETER num_ctx 32768
SYSTEM """You are an expert AI engineering assistant. Output your chain-of-thought inside ... tags before answering."""