DeepSeek R1 put open-weight reasoningmodels on the map — but there's a catch that trips up everyone trying to run it at home: the headline model is gigantic. The honest version of "run DeepSeek locally" in 2026 is about running the right distill for your hardware, and reaching for the full model only when a task genuinely needs it. This guide draws that line clearly.
Step 1 — Understand what you can and can't run
Full DeepSeek R1is roughly a 671B-parameter mixture-of-experts model. Even aggressively quantized it needs hundreds of gigabytes of memory — that's a multi-GPU server, not a desktop. So what people actually run locally are the R1 distills: smaller dense models (built on Qwen and Llama) fine-tuned on R1's reasoning traces. You keep much of the step-by-step reasoning at a size your GPU can hold.
| Model | Params | ~VRAM (Q4_K_M) | Runs on |
|---|---|---|---|
| R1-Distill 1.5B | 1.5B | ~1.5 GB | Almost anything |
| R1-Distill 7B | 7B | ~5 GB | 8 GB GPUs — the easy default |
| R1-Distill 14B | 14B | ~10 GB | 12 GB+ GPUs |
| R1-Distill 32B | 32B | ~20 GB | 24 GB card (3090/4090) |
| R1-Distill 70B | 70B | ~40 GB | 48 GB card / dual-GPU / big Mac |
| Full DeepSeek R1 | ~671B MoE | 100s of GB | Multi-GPU server (not local) |
Find the biggest distill your card can hold: GPUs for the 7B distill · start from your GPU and see all of them.
Step 2 — Run a distill
Same on-ramps as any open model. Ollama is the quickest; LM Studio or Jan give you a GUI. Start with the 7B distill unless you know you have headroom.
# Run a DeepSeek R1 distill with Ollama ollama pull deepseek-r1:7b # fits ~8 GB GPUs ollama run deepseek-r1:7b # Bigger distill if you have the VRAM: ollama pull deepseek-r1:32b # ~20 GB
For control over the exact GGUF and flags, use llama.cpp. Reasoning models emit long chains of thought, so give them a generous --ctx-size if you can.
# llama.cpp — a specific distill GGUF llama-server \ --hf-repo bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF \ --hf-file DeepSeek-R1-Distill-Qwen-7B-Q4_K_M.gguf \ --n-gpu-layers 999 \ --ctx-size 16384 \ --host 0.0.0.0 --port 8080
One reasoning-specific tip: these models "think" in a long preamble before answering, which burns tokens and time. Budget for it, and read GGUF quantization explained before going below Q4 — reasoning quality degrades faster than chat quality at low bit-depths.
Step 3 — Get the full model without buying a server
Here's the move that makes DeepSeek practical: run a distill locally for the bulk of your work, and fail over to the full DeepSeek model in the cloud only for the few requests that truly need it. You get local-and-free for the 90% case and frontier reasoning for the 10%, without owning a rack.
Wide Area Intelligenceis built for exactly this split. It's a gateway (not a runtime) that sits in front of your local DeepSeek distill and routes intelligently:
Local first, free
Cloud failover for the full model
Reachable from anywhere
wai_sk_… keys mean the homelab GPU is usable from your laptop, with no port forwarding.To be clear about positioning: Wide Area Intelligence doesn't host the 671B model on your behalf the way a cloud provider does — it makes your GPU serve the distill and hands off to real cloud only when needed, behind one endpoint.
Putting it together
Skip the fantasy of running full R1 on a gaming PC. Pick the largest R1 distill your GPU holds (the GPU-for-model tool tells you which), run it with Ollama or llama.cpp, and if you want the full model on tap for the hard requests — plus that GPU reachable from anywhere — set up Wide Area Intelligence → Free for up to two nodes.
Related: How to run Qwen locally, How to run Llama 3 locally, and Best OpenRouter alternatives.