How to Install DeepSeek on Windows

About this video

David shows two ways to run the DeepSeek R1 model locally, totally disconnected from the internet, to address privacy concerns. For non-technical users he uses LM Studio: download the correct version — on Windows check system type via MSINFO32, choosing the ARM 64 version if ARM appears, otherwise the other version — install it, and open it. Models are found via the search icon, which searches Hugging Face. The DeepSeek R1 results are distilled models, one distilled into Qwen 7B and one into Llama 8B. He explains distillation with a library analogy: the full model is too big, so distilled models are smaller summarized collections; 7B and 8B are parameter counts, and more parameters need more memory and processing power. He recommends starting with Llama unless multiple languages are needed, in which case try Qwen. He disconnects from his network and chats locally, seeing the chain-of-thought reasoning, though the model wrongly counts two R’s in Cranberry. For technical users he uses Ollama: download and install it, confirm it is running in the taskbar, open a command prompt with CMD, find DeepSeek R1 on the Ollama site, and run the command for the 8 billion parameter Llama model. The command line becomes the chat interface, showing reasoning between think tags and markdown in answers, and works without WiFi. He then compares the local distilled model with the full-sized DeepSeek R1 on the web: the full model correctly counts three R’s in Cranberry; both rewrite an email well with the full model firmer; on a math problem both technically get the correct answer but only the larger model gets the final answer; on sibling advice the larger model gives detailed step-by-step suggestions while the smaller is high level; both fail a five-hat logic riddle, with the local model thinking for almost 19 minutes; and the distilled model’s music recommendations contain hallucinations, with only Welcome to the Black Parade being a real song in its list. He suggests trying larger 14 billion or 32 billion parameter models on a beefier machine and experimenting to find the right distillation.

Key takeaways

• Run DeepSeek R1 locally and offline with LM Studio for a graphical interface or Ollama for the command line
• The downloadable models are distilled versions (Qwen 7B, Llama 8B), not the full R1; parameters indicate size, complexity, and hardware demands
• On Windows, check MSINFO32 system type to pick the correct LM Studio download, including the ARM 64 version if applicable
• In tests the full web model outperformed the local distilled model on counting, advice detail, and music recommendations, where the small model hallucinated songs
• If results are weak, try a larger distillation such as 14B or 32B on stronger hardware and experiment across models

Episode details: published February 1, 2025 · 8:46 long. Summary written by Self Growth Videos from the episode itself. Business results and any income figures mentioned are Kevin Stratvert’s own claims, shared for education — they are not typical or guaranteed.

← Previous: How to Use DeepSeek · Next: 7 Grammarly Tricks You Need to Start Using Today! →

All Kevin Stratvert videos · Kevin Stratvert profile · Kevin Stratvert on YouTube