the ONLY way to run Deepseek...

About this video

Chuck argues the only way he recommends using DeepSeek R1 is locally, not through its app or website, because anything typed into the hosted service is stored on servers in China and subject to Chinese cybersecurity laws that give authorities broad access powers — a concern he extends to hosted AI generally, but treats as sharper here. He recaps why DeepSeek rattled the industry: according to DeepSeek (his caveat: take it with a grain of salt), it was trained for less than $6 million using 2,000 Nvidia H800 GPUs, versus his comparison figure of OpenAI spending over $100 million with 10,000-plus top-tier GPUs, suggesting clever post-training engineering can beat raw compute. Because the models are open source, viewers can run them at home via LM Studio (a friendly GUI that checks whether your GPU can fully or partially offload a model) or Ollama (CLI-only). Sizes run from 1.5 billion parameters up to the full 671B model, which he says even his dual-4090 AI server cannot run; most people will live in the 1.5B–32B range depending on VRAM and RAM. He then verifies safety empirically: a PowerShell script monitors Ollama’s connections and shows it only listening locally on port 11434, with external connections appearing only when downloading a new model. For maximum isolation, his preferred setup runs Ollama inside a hardened Docker container — GPU access via the Nvidia container toolkit, capped resources, dropped privileges, and a read-only filesystem — while noting Mac users are left out because Docker cannot access M-series GPUs.

Key takeaways

• Hosted DeepSeek sends your data to servers in China; running the open-source model locally keeps prompts on your own machine.
• LM Studio is the beginner-friendly GUI option; Ollama is the simple CLI option — both can download and run DeepSeek R1 in minutes.
• Model size is the hardware gate: the full model is 671 billion parameters, but distilled versions go down to 1.5B, which most computers can attempt.
• He verified with a connection-monitoring script that Ollama makes no external connections while chatting — only when downloading models.
• For the strongest isolation, run Ollama in Docker with GPU passthrough, resource caps, dropped privileges, and a read-only filesystem.

Episode details: published January 31, 2025 · 11:59 long. Summary written by Self Growth Videos from the episode itself. Business results and any income figures mentioned are NetworkChuck’s own claims, shared for education — they are not typical or guaranteed.

← Previous: the hacker’s roadmap (how to get started in IT in 2025) · Next: Your Remote Desktop SUCKS!! Try this instead (FREE + Open Source) →

All NetworkChuck videos · NetworkChuck profile · NetworkChuck on YouTube