Deploy DeepSeek R1 Locally | 0xClaw

Why open weights changed the conversation

In early 2025, DeepSeek R1 moved local-model deployment from novelty to practical option. It is an open-weights reasoning model that holds up reasonably well on coding and logic tasks against far more expensive proprietary systems.

The more important shift was access. When the weights are downloadable, teams no longer have to assume that every sensitive prompt, codebase, or internal document must pass through a third-party API. That single property changes who can use a strong reasoning model and where.

Running DeepSeek R1 with Ollama

Modern inference tooling makes local deployment straightforward. With Ollama, the whole setup on a machine you control is:

# 1. Install the Ollama inference engine
curl -fsSL https://ollama.com/install.sh | sh

# 2. Pull and run a distilled DeepSeek R1 model
# (Pick the parameter size that fits your hardware)
ollama run deepseek-r1:14b

Once running, Ollama serves an OpenAI-compatible API at http://localhost:11434/v1. That compatibility layer matters: anything that can talk to the OpenAI Chat Completions format can now talk to the model on your own hardware.

If you would rather not run inference yourself, DeepSeek also offers a hosted API with the same model family (deepseek-chat, deepseek-reasoner). The trade-off is straightforward: the hosted API removes hardware work but sends your prompts to a third party; the local endpoint keeps every token on your machine.

Connecting the model to 0xClaw

0xClaw works against OpenAI-compatible model endpoints through its BYOK (bring your own key) settings, with two relevant paths:

1. DeepSeek API preset: paste a DeepSeek API key and 0xClaw talks to the hosted deepseek-reasoner endpoint directly.
2. Custom (OpenAI-compatible) provider: point 0xClaw at any OpenAI-compatible base URL — including the Ollama endpoint on your own machine — and supply the key or token your endpoint expects.

The second option is what makes a fully local setup possible: the model runs on your hardware, 0xClaw runs on your hardware, and prompts and outputs never leave your environment.

What this unlocks for security testing

This is where a locally deployed reasoning model becomes more than an infrastructure exercise. 0xClaw is a local AI pentest tool: it plans and executes real application security workflows — reconnaissance, vulnerability discovery, verification — and produces report-ready evidence. Pairing it with a local model endpoint means the reasoning layer and the evidence stay on your side of the network.

If that combination is new to you, two starting points:

• How a full local AI pentest workflow fits together, and what an AI pentest CLI actually does.
• Where model-level scanning (garak-style) differs from application-layer pentesting, if you are deciding which layer your testing budget should cover.

You can download 0xClaw for Mac, Linux, and Windows and point it at the endpoint you just stood up.

The takeaway

Open-weights models like DeepSeek R1 removed the assumption that strong reasoning has to come from a remote API. Once both the model and the tooling using it run on infrastructure you control, data boundaries stop being a selling point you argue about and become a property you can verify.