~/runyourllm

Run DeepSeek Locally: Real Hardware Numbers by Model Size

By , Ready Utilities

Last updated: September 2026

Run DeepSeek locally and you'll immediately hit a confusing problem: search for its hardware requirements and you'll find wildly different answers. One source says a 7B model needs 24GB of VRAM, another says 5.5GB. Both can't be right for the same model at the same quantization. Here's what's actually going on, and the real numbers by size.

Which "DeepSeek" do you mean?

This matters more than it sounds. "DeepSeek" on its own usually refers to the full DeepSeek-R1 or DeepSeek-V3-class models, which are enormous mixture-of-experts models built for data-center hardware, not a home GPU. What most people running DeepSeek locally actually use are the R1-Distill models: smaller, dense models distilled from R1's reasoning ability onto existing open architectures. A DeepSeek-R1-Distill-14B, for example, is trained on top of Qwen2.5-14B's actual architecture, not a separate DeepSeek design. That's the version this guide covers.

Real numbers by size

ModelQ4 VRAMRealistic minimum GPU
R1-Distill-Qwen-1.5B~2.3 GBAny GPU with 4GB+
R1-Distill-Llama-8B~6 GBRTX 4060 8GB
R1-Distill-Qwen-14B~9.6 GBRTX 4070 12GB
R1-Distill-Qwen-32B~22.9 GBRTX 4090 24GB
R1-Distill-Llama-70B~49.1 GBMac Studio 64GB (unified memory)

These figures come from this site's own calculator formula, not a static table copied from elsewhere: parameter count times bytes-per-weight at Q4_K_M, plus real per-token context memory pulled from each model's actual architecture (the R1-Distill-14B and R1-Distill-32B rows use the exact same architecture data as the Qwen 14B and Qwen 32B entries in the hardware calculator, since that's genuinely what they're built on).

Why other sources disagree so much

A few reasons show up repeatedly when different guides give different numbers for the same model:

  • Some are quoting the full model, not a distill. "24GB minimum" figures often describe running something closer to full R1, not the 7B or 14B distills most people actually run at home.
  • Some don't specify quantization. FP16 and Q4 numbers for the same model differ by roughly 3.5x. A table without a quantization column is missing the single most important variable.
  • Some are simply old. Model releases and community consensus on "recommended" hardware shift fast in this space; a guide written months ago may describe an earlier distill lineup.

Running it

The fastest path is the same as any other model: install Ollama, then pull the size that fits your hardware.

ollama run deepseek-r1:14b

Swap the size for whatever fits per the table above. See the full beginner guide if you haven't installed Ollama yet.

VRAM figures are computed with this site's own documented formula (see the hardware calculator's methodology section), using real architecture data for the R1-Distill-Qwen models pulled from HuggingFace configuration files. DeepSeek's own model card confirms the R1-Distill-Qwen-14B and -32B models use the Qwen2.5 architecture directly. This page gets rechecked as new DeepSeek distill sizes are released.

Sources

  1. LLM Hardware

Frequently asked questions

Is DeepSeek-R1-Distill the same thing as full DeepSeek?

Not exactly. It's a smaller model trained to imitate R1's reasoning style, built on an existing open architecture (Qwen or Llama, depending on size) rather than DeepSeek's own much larger design. It's what almost everyone running DeepSeek locally actually uses.

Can I run the full DeepSeek-R1 at home?

Not realistically. The full model is built for data-center hardware, with memory requirements far beyond any consumer GPU or Mac. The distilled versions in the table above are the practical local option.

Which DeepSeek size should I start with?

Match it to your hardware using the table above, or run your exact numbers through the calculator. If you're unsure, the 8B distill is a reasonable starting point on most modern GPUs with 8GB or more of VRAM.