Can DeepSeek V4 run on your own computer
Following this course, Xiaolin got deepseek-r1:8b running on his own 16 GB graphics card. A few days later, he saw news saying DeepSeek had already released V4.1, and articles online claiming "a 8 GB graphics card can run V4.1". He opened Ollama, searched for deepseek, and found deepseek-v4.1-flash, but the tag was cloud.
Let's assess first:
Quiz timeXiao Lin runs ollama run deepseek-v4.1-flash:cloud — where are his prompts processed?
The weights of DeepSeek V4 and V4.1 are open source, but the models are too large for a home computer. LM Studio officially says V4-Flash needs at least 156 GB of memory; the V4 series on Ollama only has cloud versions. To run DeepSeek locally on your own computer, what's currently usable is still the small size of deepseek-r1.
What you'll be able to do after this lesson
- Explain which models the DeepSeek V4 series includes and how big each one is
- Distinguish between
:cloudmodels and local models on Ollama - Decide what your own computer should run
Where DeepSeek is now
According to the timeline of DeepSeek's official News (no R2):
| Date | Release | Officially stated scale |
|---|---|---|
| 2026-04-24 | V4 preview, open source | V4-Pro: 1.6T total parameters, 49B activated; V4-Flash: 284B total parameters, 13B activated |
| 2026-08-13 | V4-Pro official release | Reasoning intensity split into low / high / max |
| 2026-09-10 | V4.1-Flash, natively understands images | MoE with 552B parameters; 8B activated for input, 16B activated for output |
| 2026-10-17 | V4.1-Flash-1017, open-sourced on Hugging Face and ModelScope under the MIT license | Same as above |
When V4.1-Flash was released, V4-Flash on the official API had already been taken offline. And deepseek-r1, which this course has been using all along, shows "1 year ago" in the Ollama model library.
There are two versions of the parameter count for V4.1-Flash — take note: DeepSeek's news says 552B, while the weight files of V4.1-Flash-1017 on Hugging Face count 763,205,315,794 parameters (about 763B), and the Ollama page also says 763B. Where the difference comes from hasn't been explained officially.
- I know the latest from DeepSeek is the V4 / V4.1 series, not R2
Why are all the V4 models on Ollama cloud versions
Searching for deepseek on Ollama, both models in the V4 series are only tagged cloud:
| Model | Available tags | Notes |
|---|---|---|
deepseek-v4.1-flash | :cloud, :1017-cloud | 1M context, text and image input |
deepseek-v4-pro | Cloud version only | 1M context |
To run it:
ollama run deepseek-v4.1-flash:cloudWhat you should see: no multi-hundred-GB file download, because Ollama's official documentation says cloud models don't need to be downloaded. The prerequisite is logging in to your Ollama account first (ollama signin, or logging in on the settings page of the Mac or Windows app); the model page is priced per million tokens, and you pay for what you use.
The biggest difference between the cloud and local versions is privacy:
deepseek-r1:8b (local) | deepseek-v4.1-flash:cloud (cloud) | |
|---|---|---|
| Where inference happens | Your graphics card | Ollama's cloud service |
| Your questions | Stay on your machine | Pass through Ollama's service |
| Hardware needed | VRAM that fits a 5.2 GB file | No good graphics card needed |
| Account needed? | No | Must log in to an Ollama account |
Ollama officially says it processes cloud questions and answers to produce results and doesn't use them to train models. If you're deploying locally precisely to keep data from leaving your computer, don't use :cloud models; after you turn off the cloud feature using the method in the FAQ, :cloud models will no longer work either.
- I can tell that when a model name has
:cloud, it doesn't run on my computer
If you really want to run V4 locally, what hardware do you need
The only numbers that can be found in official materials are these:
Look it upI want to run it locally...
Click one to see.
Claims online like "a 8 GB graphics card runs V4.1" have no basis in any of the official materials above. Even for V4-Flash with fewer parameters (284B), LM Studio's official lower bound is 156 GB of memory.
- I know V4-Flash needs at least 156 GB of memory to run locally
How to choose for a home computer
Putting the above together:
Data must stay on your machine: use the small sizes of deepseek-r1, and follow which model size to choose to pick one that fits entirely in VRAM.
You want the latest V4.1 and can accept sending prompts to the cloud: use
deepseek-v4.1-flash:cloud, or just use DeepSeek's official web page and API.You have a machine with 156 GB of RAM or more: you can try running V4-Flash locally following LM Studio's blog post. This course has not tested it.
I have chosen one path based on my own situation
Checkpoint
- I can state the total parameters of V4-Flash and V4-Pro (284B, 1.6T)
- I know that on Ollama the V4 series only comes in the
:cloudversion, which requires logging into an account and is billed per token - I know that when using the
:cloudmodel, prompts pass through Ollama's service - I know that what a 16 GB GPU can run locally is the small sizes of deepseek-r1
Common misconceptions
Assuming "open source" means "I can run it on my own computer." The weights of V4 and V4.1 are indeed open source and anyone can download them, but running V4-Flash locally requires at least 156 GB of RAM. Open source means the weights are available, not that the hardware requirements are low.
Assuming everything running in Ollama is a local model. Models whose names end with :cloud run on Ollama's cloud and are not downloaded to your machine. The earlier statement in this course, "when running locally they cannot see your prompts," only holds for local models.
Assuming that small activated parameters mean VRAM is enough. The official figures for V4-Flash are 284B total parameters and 13B activated. LM Studio's minimum for this model is at least 156 GB of system RAM, not a number estimated from 13B.
Next section
With a locally runnable model chosen, use it to build your own knowledge base: Build Your Own Knowledge Base.