DeepSeek’s R1 reasoning model made headlines in January 2025 as an open model that “thinks” step by step before answering. Its smaller distilled versions remain some of the easiest reasoning models to run on a normal computer. In 2026, DeepSeek has also released the much larger V4 family. This guide explains which DeepSeek model fits your hardware, how to run it with Ollama or LM Studio, and what speed to realistically expect.
Which DeepSeek Model Can You Run?
The full DeepSeek-R1 is a 671-billion-parameter Mixture-of-Experts model, far too big for normal PCs. The practical choices are the R1 distills: smaller Qwen- and Llama-based models trained to imitate R1’s reasoning.
| Model (Ollama tag) | Approx. size at 4-bit | Memory to run comfortably |
|---|---|---|
R1 distill 1.5B (deepseek-r1:1.5b) |
~1.1GB | Any 8GB machine |
R1 distill 7B / 8B (deepseek-r1:7b, :8b) |
~4.5–5GB | 16GB RAM or an 8GB GPU |
R1 distill 14B (deepseek-r1:14b) |
~9GB | 16GB GPU or 24GB+ Mac |
R1 distill 32B (deepseek-r1:32b) |
~19–20GB | 24GB GPU or 32GB+ Mac |
R1 distill 70B (deepseek-r1:70b) |
~42–43GB | 64GB+ Mac, Strix Halo or 2× 24GB GPUs |
| Full R1 671B | ~130GB even at extreme ~1.6-bit quantization | Workstation or server-class memory |
DeepSeek’s May 2025 update (R1-0528) also produced a newer 8B distill based on Qwen3, which is worth trying if you’re limited to 16GB. Use the VRAM Calculator to check a size and context length for your machine.
How to Run It with Ollama
- Install Ollama from ollama.com. The current release in September 2026 is the v0.33 series.
- Open a terminal and run a size that fits your memory, for example:
ollama run deepseek-r1:14b - Ollama downloads the model and opens a chat. The model shows its reasoning (“thinking”) before the final answer.
Prefer a graphical app? LM Studio lets you search for “DeepSeek R1” and download a GGUF build, or an MLX build on Apple Silicon. See our comparison of local LLM runners.
What Speed to Expect (Estimates)
Generation speed is mostly limited by memory bandwidth. Using the rule of thumb tokens/sec ≈ bandwidth × 0.6–0.75 ÷ model size in GB:
| Hardware | 14B distill (~9GB) | 32B distill (~20GB) |
|---|---|---|
| Mac mini M5 Pro (~307GB/s) | ~20–26 tok/s | ~9–12 tok/s (needs 48GB+ config) |
| RTX 5060 Ti 16GB (448GB/s) | ~30–37 tok/s | Doesn’t fit in 16GB |
| RTX 4090 24GB (1,008GB/s) | ~67–84 tok/s | ~30–38 tok/s |
| Mac Studio M5 Max (614GB/s) | ~41–51 tok/s | ~18–23 tok/s |
These are estimates from official specs, not our own measurements. Keep in mind that reasoning models write long “thinking” sections, so even at a good speed an answer can take a while.
Newer Options: DeepSeek V4 and Other Reasoning Models
DeepSeek released DeepSeek-V4 on April 24, 2026, in two open-weight versions under the MIT licence:
- V4-Flash: ~284–291B total parameters, with about 13B active per token.
- V4-Pro: ~1.6T total parameters, with about 49B active per token.
Both use a Mixture-of-Experts design with a 1-million-token context window. Even V4-Flash needs roughly 170GB+ of memory at 4-bit, putting it in Mac Studio M5 Ultra or multi-GPU territory. See our large-model sizing guide. On a normal PC, newer mid-size open models such as Qwen3.8-27B, Gemma 4 and gpt-oss-20b are strong alternatives to the R1 distills.
Tips for Better Results
- Leave room for context: reasoning chains are long, so budget extra memory for the KV cache.
- Pick 4-bit (Q4_K_M) first: it’s the usual sweet spot between size and quality. See our quantization guide.
- Keep the whole model in fast memory: if part of it spills into system RAM, speed drops sharply.
FAQ
Is the local DeepSeek R1 the same as the DeepSeek app?
No. The small models you run locally are distills, which are much smaller than the full R1 or V4 models behind DeepSeek’s app and API. What those full models cost to rent, and what they would actually need to run locally, is worked through in DeepSeek API pricing vs running it locally.
Is running DeepSeek locally private?
Yes. Once downloaded, the model runs entirely on your computer and your prompts don’t leave your machine.
Sources
DeepSeek on Hugging Face, DeepSeek V4 release notes, Ollama DeepSeek-R1 library.
Updated September 11, 2026: added DeepSeek V4, current Ollama details and clearly labelled speed estimates in place of earlier benchmark claims.
Work out what your hardware can actually run
Free calculators and directories, sized from real model weights — before you download gigabytes or order a card:
