Deploy DeepSeek‑R1 Locally with Ollama and Connect It to Webman AI

Overview

DeepSeek‑R1 is an open‑source large language model released under the MIT license, permitting unrestricted commercial use. Its benchmark results are comparable to leading proprietary models.

Why Deploy Locally?

Public DeepSeek endpoints have suffered intermittent outages due to external attacks. Running the model on a local GPU guarantees reliable, continuous access.

Prerequisites

A computer equipped with a GPU. The amount of VRAM determines which model size can be run efficiently.

Install Ollama

Download the installer from https://ollama.com/download and follow the platform‑specific installation steps.

Select Model Version

Choose a model size that matches your GPU’s VRAM. Use the corresponding ollama run command to launch the model.

1.5B model – ~4 GB VRAM (e.g., GTX 1050) – ollama run deepseek-r1:1.5b 8B model – 8‑10 GB VRAM (e.g., GTX 1660) – ollama run deepseek-r1:8b 14B model – ≥12 GB VRAM, 16 GB recommended (e.g., RTX 3060) – ollama run deepseek-r1:14b 32B model – ≥16 GB VRAM, 21 GB recommended (e.g., RTX 3060) – ollama run deepseek-r1:32b 70B model – ≥24 GB VRAM, 40 GB recommended (e.g., RTX 3090 / RTX 4090) – ollama run deepseek-r1:70b 671B model – 1.34 TB VRAM (e.g., 16 × NVIDIA A100 80GB) –

ollama run deepseek-r1:671b

If the GPU does not meet the VRAM requirement, Ollama will fall back to a CPU‑plus‑GPU mode, which is functional but very slow (approximately 1–2 characters per second).

Run a Conversation

After Ollama is installed, start a chat session from the command line, substituting the version you installed:

ollama run deepseek-r1:14b

Optional: Integrate with Webman AI

If you have a Webman AI instance, you can register the locally‑deployed DeepSeek‑R1 model to obtain a web UI and improve user interaction.

Integration Steps

Open the Webman AI admin console.

Select “Add Model” and enter the external model name (e.g., deepseek-r1).

Create a role for the model.

Test the connection to verify that the model responds.