DeepSeek's reasoning models are free and open — but the full one needs a datacenter. The good news: the distilled versions run great on a normal PC. Here's how to run DeepSeek locally, and which size fits your hardware.
DeepSeek's R1 reasoning models are free, open, and genuinely impressive — which is why 'how to run DeepSeek locally' is one of the most-searched local-AI questions. There's one catch worth knowing up front: the full DeepSeek-R1 is 671 billion parameters and needs a datacenter to run. But that's not the version most people run at home. The distilled versions — from 1.5B up to 70B — pack much of R1's reasoning ability into sizes that fit a normal PC, and they run with a single command. Here's how to run DeepSeek locally, and how to pick the size that fits your hardware.
Run it with one command
The simplest route is Ollama. Install it, then run one line and it downloads and starts a distilled DeepSeek-R1 model. Pick the size that matches your hardware — here's the pattern:
# install Ollama, then pick a size that fits your card:
ollama run deepseek-r1:8b # ~8GB VRAM
ollama run deepseek-r1:14b # ~16GB VRAM
ollama run deepseek-r1:32b # ~24GB VRAM
# prefer a graphical app? LM Studio runs the same models.
That's it — the first run downloads the model, then you're chatting with a reasoning model that thinks step-by-step before answering. If you'd rather click than type, LM Studio runs the same distilled DeepSeek models through a friendly interface.
The distilled DeepSeek-R1 models bring step-by-step reasoning to hardware you already own. · Unsplash
Which size fits your hardware
Match the model to your VRAM. On an [8GB card](/best-local-llm-8gb-vram-2026), run the 7B or 8B distill — light and quick. On a [16GB card](/best-local-llm-16gb-vram-2026), the 14B distill is the sweet spot. On a [24GB card](/best-gpu-for-qwen3-32b-local-2026), the 32B distill is the most capable single-GPU option and noticeably stronger at reasoning. The 70B distill needs more memory than one consumer card — think dual GPUs or a big unified-memory Mac. The smaller distills are less capable than the full R1, but they retain a lot of its reasoning style, and for local use they're excellent. Size yours precisely in the VRAM calculator.
Quick answers
Can I run DeepSeek on my own computer?
Yes — the distilled DeepSeek-R1 models (1.5B to 70B parameters) run on normal PCs, and you can start one with a single Ollama command like 'ollama run deepseek-r1:8b'. The full DeepSeek-R1 (671B parameters) needs datacenter hardware and isn't runnable at home, but the distilled versions keep much of its step-by-step reasoning ability and fit consumer graphics cards. Pick the size that matches your VRAM and you're running DeepSeek locally, free and private.
Which DeepSeek version can I run on my GPU?
It depends on your VRAM: an 8GB card runs the 7B or 8B distill, a 16GB card runs the 14B, and a 24GB card runs the 32B — the most capable single-GPU option. The 70B distill needs more than one consumer card (dual GPUs or a large unified-memory Mac). The full 671B model is datacenter-only. For local use, the 14B and 32B distills hit the best balance of reasoning quality and hardware you can actually own.
Is running DeepSeek locally free?
Yes — DeepSeek's models are open and free to download, and the software to run them (Ollama or LM Studio) is free too. Once you have a capable graphics card, running DeepSeek locally costs only electricity, with no subscription or API fees. It's also private and offline: your prompts never leave your computer. That's a key difference from using DeepSeek through a cloud service, where your data goes to a server.
Running DeepSeek locally is one command away with the distilled models — pick a size for your card and go. For the full model's real datacenter cost, see our DeepSeek-R1 cost breakdown; new to local AI, start with the beginner's guide.
