# How to run DeepSeek locally: Ollama commands and model sizes

> Small DeepSeek-R1 models run on a laptop with one Ollama command. The 763B V4.1 Flash needs a server, and on Ollama it is cloud-only.

URL: https://claudegpt.si/news/deepseek-how-to-run-locally  
Published: 2026-11-06T10:54:27.595Z  
Author: ClaudeGPT Si Newsroom  
Section: Guides  
Topics: DeepSeek, Ollama, Open-weight models, Hugging Face  
Publisher: ClaudeGPT Si

## Key takeaways

- The easiest way to run DeepSeek locally is Ollama: install it, then run ollama run deepseek-r1:8b, a 5.2 GB download.
- Only the 671b tag (404 GB) is the full DeepSeek-R1; smaller tags are distilled Qwen and Llama models.
- DeepSeek V4.1 Flash has MIT-licensed weights at 763B parameters, and Ollama lists it only with cloud tags.

You can run DeepSeek locally, on your own computer, but only the smaller DeepSeek-R1 models fit on ordinary hardware. The easiest way is [Ollama](https://ollama.com/library/deepseek-r1): install it, then run `ollama run deepseek-r1:8b`. DeepSeek's current flagship, V4.1 Flash, has open weights too, but at 763B parameters it needs a server, not a laptop.

This guide is for developers and curious users who want DeepSeek running offline, with no account and no API bill. By the end you will have a model answering in your terminal and a local API on your machine. Facts come from the Ollama library, Hugging Face model cards and DeepSeek's GitHub page, as of November 6, 2026. We did not run the commands ourselves, so check the sizes against your hardware before you download.

## Which DeepSeek model can you run locally?

Ollama's [DeepSeek-R1 page](https://ollama.com/library/deepseek-r1) lists the sizes below. The 128K context applies to every tag except the full model, which has 160K. The download size is a guide to the memory you need: the model has to fit in your RAM, or your GPU's VRAM, to run at a usable speed.

| Tag | Download | Based on |
| --- | --- | --- |
| `deepseek-r1:1.5b` | 1.1 GB | Qwen distilled model |
| `deepseek-r1:7b` | 4.7 GB | Qwen distilled model |
| `deepseek-r1:8b` (the default) | 5.2 GB | Qwen3-8B |
| `deepseek-r1:14b` | 9.0 GB | Qwen distilled model |
| `deepseek-r1:32b` | 20 GB | Qwen distilled model |
| `deepseek-r1:70b` | 43 GB | Llama distilled model |
| `deepseek-r1:671b` | 404 GB | The full DeepSeek-R1 |

A distilled model is a small model trained on the answers of a bigger one. Only the 671b tag is the real DeepSeek-R1. The others are Qwen and Llama models that DeepSeek taught to reason like it, so they are weaker than the full model. The 8B tag is DeepSeek-R1-0528-Qwen3-8B. Its [model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528) says it scores 86.0 on the AIME 2024 math test, against 76.0 for Qwen3-8B. That is DeepSeek's own number.

Newer DeepSeek models are a different story. DeepSeek V4.1 Flash is a [763B-parameter model](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) under the MIT license, and the model card gives no hardware requirements or file sizes. On Ollama it is listed only with `:cloud` tags, such as `ollama run deepseek-v4.1-flash:cloud`. That runs on Ollama's servers, not on your machine. Our [V4.1 Flash explainer](/news/deepseek-v4-1-flash-explained) covers what the model is, and [is DeepSeek open source](/news/is-deepseek-open-source) covers the licenses.

## How to run DeepSeek with Ollama

Ollama is the quickest route on macOS, Windows and Linux.

1. Download Ollama from [ollama.com/download](https://ollama.com/download) and open it. On Linux, if the server isn't running, start it with `ollama serve`.
2. Pick a size from the table that fits your memory. If you are unsure, start with 8b.
3. Download and start a chat:

```shell
ollama run deepseek-r1:8b
```

The first run downloads the model, then opens a chat in the terminal. Type `/bye` to exit. To download without chatting, use `ollama pull deepseek-r1:8b`.

4. To update an older copy, run `ollama pull deepseek-r1`. Ollama says the model had a minor upgrade to DeepSeek-R1-0528 for the 8B and 671B tags.

## How to use your local DeepSeek from code

Ollama's [quickstart](https://docs.ollama.com/quickstart) says the local server listens on port 11434 and needs no API key. A request looks like this:

```shell
curl http://localhost:11434/api/chat \
  -H "Content-Type: application/json" \
  -d '{"model": "deepseek-r1:8b", "messages": [{"role": "user", "content": "Say hello in one sentence."}], "stream": false}'
```

Ollama also serves an OpenAI-compatible endpoint at `/v1/chat/completions` and an Anthropic-style one at `/v1/messages`, so most tools that accept a custom base URL can point at it. If you would rather use DeepSeek's hosted API, see [how to get a DeepSeek API key](/news/deepseek-how-to-get-api-key).

## How to run DeepSeek with vLLM or SGLang

If you have GPUs and want a server, DeepSeek's [R1 repository](https://github.com/deepseek-ai/DeepSeek-R1) gives these commands for a distilled model, split across two GPUs:

```shell
vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --tensor-parallel-size 2 --max-model-len 32768 --enforce-eager
python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --trust-remote-code --tp 2
```

The V4.1 Flash model card lists `vllm serve "deepseek-ai/DeepSeek-V4.1-Flash-1017"` and an SGLang launch command too. It also lists `docker model run hf.co/deepseek-ai/DeepSeek-V4.1-Flash-1017` for Docker Model Runner. The card says the repository's `inference` folder holds the weight conversion and local inference code. There is no Jinja chat template. DeepSeek ships a Python encoder, `encoding.py`, instead.

## Settings that DeepSeek recommends

- For R1 distilled models, DeepSeek's README says to set the temperature between 0.5 and 0.7, with 0.6 recommended, to avoid repetition or incoherent output.
- The same README advises against a system prompt and says to put all instructions in the user prompt. The R1-0528 card says that version does support a system prompt and no longer needs a forced `<think>` at the start of the output.
- For math, DeepSeek suggests adding "Please reason step by step, and put your final answer within \boxed{}."
- For V4.1 Flash, the model card recommends temperature 1.0, top_p 1.0 and a `max_tokens` of at least 256K.

## Common problems

- **The download is huge or the model is slow.** Pick a smaller tag. If the model doesn't fit in memory, it runs far slower.
- **Answers loop or ramble.** Lower the temperature to around 0.6 for R1 models.
- **The model shows its reasoning at length.** R1 models are reasoning models and think before they answer. That is expected.
- **You expected the full DeepSeek.** Only `deepseek-r1:671b` is the full R1. Smaller tags are distilled Qwen and Llama models.
- **A guide tells you to use odd flags.** Check `ollama --help` first. Some guides online list options that Ollama doesn't document.

## What to do next

If your machine is too small, the hosted route costs little. See [DeepSeek pricing](/news/is-deepseek-free) for the API price list, and [DeepSeek not working](/news/deepseek-not-working-fixes) if the app or API fails.

## Frequently asked questions

### Can DeepSeek be run locally?

Yes. DeepSeek-R1 and its distilled versions run locally through Ollama, vLLM or SGLang. The full models are very large: the full R1 is a 404 GB download on Ollama.

### How do I run DeepSeek locally on Windows or Mac?

Download Ollama from ollama.com/download for Windows or macOS, open it, then run ollama run deepseek-r1:8b in a terminal.

### Can I run DeepSeek V4.1 Flash locally?

The weights are public under the MIT license on Hugging Face, but the model has 763B parameters and the model card gives no hardware requirements. Ollama lists it only with cloud tags.

### Is running DeepSeek locally free?

Ollama and the DeepSeek-R1 weights are free to download, and the local server needs no API key. You pay for your own hardware and electricity.

## Sources

- [deepseek-r1 on Ollama](https://ollama.com/library/deepseek-r1), Ollama
- [deepseek-v4.1-flash on Ollama](https://ollama.com/library/deepseek-v4.1-flash), Ollama
- [Ollama quickstart](https://docs.ollama.com/quickstart), Ollama
- [DeepSeek-R1-0528 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528), Hugging Face
- [DeepSeek-V4.1-Flash-1017 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017), Hugging Face
- [DeepSeek-R1 repository](https://github.com/deepseek-ai/DeepSeek-R1), DeepSeek