DeepSeek R1: democratizing AI

On 8 February 2025 Ignacio Delgado ran a hands-on workshop at the hackerspace about DeepSeek R1, the first open language model with capabilities comparable to OpenAI’s commercial models. The material is in the workshop repo on GitHub.

Why it matters

We started with the context: AI’s “Sputnik moment”. An open model at that level breaks the monopoly big tech had, and changes who can use advanced AI and how.

The technical part

• How the model is built and its variants, from 1.5B to 671B parameters.
• What hardware you need for each one.
• Alternatives to CUDA and ways to optimise it.
• Running it locally with Ollama, which was the base for everything hands-on.

Tools we tried

• Roo-Code: analysing and writing code from your editor.
• Open WebUI: a ChatGPT-style web interface for your local models.
• Chat with PDF-RAG: asking questions about your documents.
• Page Assist: the model inside your browser.
• n8n: automations.

The debate

What might change: more open and efficient AI, less dependence on big data centres, access to advanced models for anyone, and a blow to the current business model.

And what was left open: how much it really cost to develop, whether undeclared hardware was used, whether trade restrictions were respected, whether an open model like this is sustainable, and what happens with privacy.

More on Ignacio Delgado’s blog.