Running DeepSeek models on a NAS

What are the DeepSeek-R1 distilled models?

DeepSeek trained smaller models to reason like its full R1 model, built on Qwen and Llama bases. They show their thinking before answering, which helps with maths, planning and step-by-step questions but makes responses longer. The 8B and 14B suit 16GB to 32GB systems; the 32B fits a 64GB NAS comfortably; the 70B needs a 96GB-class system (Estimated).

Can a NAS run the full DeepSeek-R1 or V3?

Not realistically. Both are mixture-of-experts models with 671B total and 37B active parameters, and around 404GB at 4-bit. Only server-class systems with very large memory, such as QNAP's h1290FX range at its 1TB ceiling, come close, and even then a GPU setup would be needed for usable speed (Estimated).

Do reasoning models suit a NAS without a GPU?

They work, but they write a lot before answering, so slower hardware feels slower still. On a CPU-only NAS, the 8B distill is the comfortable choice.

Common questions

Are DeepSeek models safe to run privately?

Running them locally means your prompts never reach DeepSeek's servers. The weights run entirely on your NAS.

Which DeepSeek model is best for a 64GB NAS?

The R1 32B distill, at about 20GB at 4-bit, fits with room for long reasoning.