CyberFreezeDev- Home
- Blog
- R&D / Labs notes
- How to run DeepSeek locally with Ollama: models and hardware
How to run DeepSeek locally with Ollama: models and hardware
Install Ollama and run deepseek-r1:8b, a 5.2 GB distilled model that fits an 8 GB GPU. Here is which size suits which hardware, how to install on each system and when local is the wrong call.
To run DeepSeek locally, install Ollama from ollama.com on Windows, macOS or Linux, then type ollama run deepseek-r1:8b. That 5.2 GB download is DeepSeek's R1-0528 reasoning model distilled into Qwen3 8B, and it fits a graphics card with 8 GB of memory. The full DeepSeek R1 has 671 billion parameters and is a 404 GB download, so on an ordinary PC you are running a smaller distilled model, not R1 itself.
Pick the size from your hardware first, then install.
Which DeepSeek model can your hardware run?
Ollama's deepseek-r1 page offers seven sizes. Six are distilled models: DeepSeek took open Qwen and Llama models and fine-tuned them on 800,000 samples generated with R1, so they reason in R1's style at a fraction of the size. Only the 671b tag is DeepSeek R1 itself.
| Ollama tag | Built on | Download | Suits |
|---|---|---|---|
| deepseek-r1:1.5b | Qwen2.5 Math 1.5B | 1.1 GB | Any recent laptop, even without a GPU |
| deepseek-r1:7b | Qwen2.5 Math 7B | 4.7 GB | An 8 GB GPU |
| deepseek-r1:8b (default) | Qwen3 8B, from R1-0528 | 5.2 GB | An 8 GB GPU |
| deepseek-r1:14b | Qwen2.5 14B | 9.0 GB | A 12 or 16 GB GPU |
| deepseek-r1:32b | Qwen2.5 32B | 20 GB | A 24 GB GPU, short context |
| deepseek-r1:70b | Llama 3.3 70B Instruct | 43 GB | 48 GB of GPU memory, one card or two |
| deepseek-r1:671b | DeepSeek R1 (37B active of 671B) | 404 GB | A multi-GPU server |
The last column is our arithmetic, not a test: the download plus room for the conversation and the runtime. Our guide to how much VRAM an LLM needs shows the sums. If the model does not fit, Ollama still runs it with part of it in system RAM, only more slowly.
Two notes on the small ones. The 1.5b and 7b were built on Qwen's maths models, so test them on your own tasks before you rely on them for writing. The 8b is the newest distill: DeepSeek distilled it from the R1-0528 update, and it is what ollama run deepseek-r1 gives you by default.
How do you install Ollama?
Windows. You need Windows 10 22H2 or newer. Download OllamaSetup.exe from ollama.com and run it; it installs in your home folder and does not need administrator rights. With an NVIDIA card, Ollama asks for driver 551.61 or newer. Models are saved in %HOMEPATH%\.ollama; if your C: drive is small, set the OLLAMA_MODELS environment variable to a folder on a bigger drive and restart Ollama.
macOS. You need macOS Sonoma (14) or newer. Open the downloaded .dmg and drag Ollama to Applications. On first start it offers to link the ollama command into /usr/local/bin. Apple M series Macs use the GPU; Intel Macs run on the CPU only.
Linux. One command installs it:
curl -fsSL https://ollama.com/install.sh | sh
Check it with ollama -v. For an NVIDIA card, nvidia-smi should list your GPU before you start; AMD cards need ROCm v7. Ollama's Linux page also shows how to run it as a systemd service so it starts with the machine.
Your first run
Open a terminal (PowerShell on Windows) and run:
ollama run deepseek-r1:8b
The first run downloads the model, then gives you a >>> prompt. Ask something that needs steps, such as a pricing calculation, and you will see the model reason first and answer after. Through the API, Ollama returns the reasoning in a separate thinking field, so an app can hide it.
Three settings are worth knowing:
- Check where it runs. In a second terminal,
ollama psshows a PROCESSOR column. "100% GPU" is what you want; a CPU/GPU split means the model did not fit and replies will be slower. - Context length. Ollama's default context is 4,096 tokens. For long documents, type
/set parameter num_ctx 8192inside the chat, or setOLLAMA_CONTEXT_LENGTHfor the server. More context uses more memory. - Temperature. DeepSeek's model card recommends 0.5 to 0.7, with 0.6 as the default. For the older distills it also advises no system prompt: put all instructions in your message.
Ollama listens only on 127.0.0.1, port 11434, so other machines cannot reach it until you change OLLAMA_HOST.
Watch out for cloud tags
Ollama now lists newer DeepSeek models, such as deepseek-v4-pro and deepseek-v4.1-flash, as cloud models: their tags end in :cloud and run on Ollama's servers, not yours. Ollama's FAQ says it does not see your prompts when you run locally, but cloud models process them on its servers. If privacy is your reason for running locally, use tags without "cloud", or set OLLAMA_NO_CLOUD=1 to switch cloud models off.
When is running DeepSeek locally the wrong choice?
- You need the full model's quality. The 671b tag needs more than 404 GB of memory. DeepSeek's newer V4 Flash is 284 billion parameters with 13 billion active, and its card points to vLLM or SGLang on a GPU server. That is a GPU server project, not a desktop one.
- Many people will use it at once. Ollama handles one request per model at a time unless you raise
OLLAMA_NUM_PARALLEL, and each parallel slot needs its own context memory. Our Ollama vs vLLM vs llama.cpp comparison covers what to use for a team. - The licence does not suit you. DeepSeek's own weights are MIT licensed. The Qwen distills come from Apache 2.0 models, but the 70b is built on Llama 3.3 and carries Meta's licence. Read it before you build a product on that size.
Running DeepSeek for a whole team
A laptop is the right place to try a distill. Putting one in front of staff means picking the size against your documents, giving it enough memory for several users and keeping it inside your network. That is what our private AI deployment work covers. Book a consultation.
