The full DeepSeek R1 is a 671-billion-parameter monster that needs a small cluster. But its distilled versions run on a single GPU — and knowing the difference saves you a fortune.
There are two very different questions hiding inside 'how much does it cost to run DeepSeek R1 locally,' and conflating them is where people waste money. The full DeepSeek R1 is a 671-billion-parameter Mixture-of-Experts model that needs roughly 400GB of memory at 4-bit — that's a small cluster or a maxed-out 512GB Mac Studio, and for almost everyone the honest answer is to rent it by the hour, not own it. But DeepSeek also released distilled models — versions from 1.5B up to 70B that inherit much of R1's reasoning ability and run on a single consumer GPU. Knowing which you actually need is the whole game.
Full model: rent, don't buy
Running the full 671B R1 locally is a datacenter project. At 4-bit it's ~400GB of weights, which means either a multi-GPU server, or a 512GB Mac Studio that runs it at usable-but-slow speeds around 17–18 tokens/sec. Owning that hardware is a five-figure decision. Renting a suitable cloud instance costs a few dollars an hour — so unless you're running the full model constantly, renting is dramatically cheaper. This is the same math as running Llama 70B or Kimi K3: for occasional use of a giant model, the cloud wins on cost by a wide margin.
The full R1 needs ~400GB of memory — a cluster or a maxed Mac Studio. For most people, renting it by the hour is the rational choice. · Unsplash
Distills: the model most people actually want
Here's the part that saves you money. DeepSeek distilled R1's reasoning into smaller models built on Qwen and Llama bases — 1.5B, 7B, 8B, 14B, 32B, and 70B. These run on ordinary hardware: the 32B distill fits a 24GB card, the 70B distill needs dual GPUs or a big Mac, and the smaller ones run on almost anything. They don't match the full model on the hardest problems, but they capture a lot of its step-by-step reasoning style for a fraction of the hardware. For the overwhelming majority of people asking 'how do I run DeepSeek R1 locally,' a distill on a card you can afford is the real answer.
Quick answers
Can I run the full DeepSeek R1 on a consumer GPU?
No — the full 671B model needs about 400GB of memory even at 4-bit, which is far beyond any consumer GPU. Your options for the full model are a multi-GPU server, a 512GB Mac Studio (which runs it slowly), or renting cloud hardware by the hour. For consumer hardware, you run one of the distilled versions instead, which are specifically designed to fit on normal GPUs while keeping much of R1's reasoning ability.
How good are the DeepSeek R1 distills?
Surprisingly good for their size — they inherit R1's reasoning-focused training, so a 32B distill reasons more carefully than a typical 32B model. They won't match the full 671B model on the hardest, most novel problems, but for everyday reasoning, coding, and analysis they're strong and they run on hardware you can actually own. The 32B distill on a 24GB card is the sweet spot most people should start with.
Is it cheaper to rent or own for DeepSeek R1?
For the full model, rent — owning the ~400GB-capable hardware is a five-figure outlay, while renting is a few dollars an hour, so unless you run it constantly the cloud is far cheaper. For a distill, owning makes sense: a 24GB card that runs the 32B distill pays for itself quickly versus repeated cloud use, and gives you privacy and no rate limits. Match the approach to model size and usage frequency.
The honest answer to running DeepSeek R1 locally: rent the full 671B model if you truly need it, but run a distill for everyday use — the 32B distill on a 24GB card is what most people actually want. Check rental rates and fit in our cost-to-run tracker and the VRAM calculator.
