You may run DeepSeek V4 locally: the weights have been on Hugging Face under the MIT license since April 22, 2026. In practice, though, V4 doesn't run on a single machine. Even the smallest model, DeepSeek-V4-Flash with 284 billion parameters, takes up about 160 GB in its repository, more than an Nvidia DGX Spark with 128 GB of memory has. V4.1-Flash is about 510 GB, V4-Pro about 865 GB. That's hardware for several GPUs or a data center. In Ollama, V4-Pro and V4.1-Flash are only available as cloud models, so your input goes to Ollama's servers. What runs locally under deepseek-r1 in Ollama is mostly small models from 2025 that were trained on DeepSeek R1's outputs, not V4. On privacy the difference is clear: through DeepSeek's app and API, your data is processed and stored in China according to the privacy policy. With your own weights on your own hardware, nothing goes to DeepSeek.
I don't run DeepSeek myself. My DGX Spark runs Qwen3.8-27B, which I use as the reference for the numbers below. The figures come from the model cards and the Hugging Face API, which I pulled on October 26, 2026. Everything I write about local AI is collected on the AI page, and the overview of all open models is in Open Source AI Models 2026 Compared.
Which DeepSeek models exist in October 2026
DeepSeek calls V4 a preview in the model card ("a preview version of DeepSeek-V4 series"). Updates followed as separate repositories, such as V4-Flash-0731 and V4-Pro-0813. V4.1-Flash is a new model that processes text and images. According to DeepSeek, the 1017 version is a minor update that finishes tasks faster and with shorter outputs. All models in the V4 line are under MIT. So you may use them commercially, modify them and pass them on, as long as you include the copyright and license notice.
How much memory DeepSeek V4 needs
The V4 weights are already stored in compressed form. According to DeepSeek, the experts are in FP4 and most of the rest in FP8. Further quantization won't make them much smaller. Repository storage according to the Hugging Face API:
The assessment is my calculation from vendor figures, not a measurement. Memory for the context comes on top. For the strongest reasoning mode, DeepSeek recommends a context window of at least 384,000 tokens. My DGX Spark reports about 119 GB as usable, my MacBook Pro M3 Max has 64 GB. V4-Flash in its published form fits on neither.
DeepSeek's own instructions assume several GPUs too. In its example, the reference implementation in the repository splits the model across four processes using model parallelism, and across eight for V4.1-Flash. Nvidia writes that up to four DGX Sparks can be linked into a cluster. I haven't tested whether V4-Flash runs usefully on two linked Sparks. What a machine for the different model classes costs is in Your Own AI Server: Hardware and Cost.
DeepSeek in Ollama: V4 cloud only
In the Ollama library, V4-Pro and V4.1-Flash only exist with the cloud tag, for example deepseek-v4-pro:cloud or deepseek-v4.1-flash:1017-cloud, with no local size listed. The page for deepseek-v4-flash says: "deepseek-v4-flash was retired on August 27, 2026" and "No models have been pushed."
A cloud model in Ollama doesn't run on your machine. Ollama describes it this way: "Run models in Ollama's cloud from your apps or terminal. No model or app download required." Your input goes to Ollama's servers. According to Ollama, the service processes it to answer the request and doesn't use it for training. So typing ollama run deepseek-v4.1-flash:cloud doesn't run DeepSeek locally, it runs it in someone else's cloud. How to set up Ollama for models that really are local is in Setting Up Local AI: Ollama to SGLang.
What "deepseek-r1" in Ollama is
Most guides on installing DeepSeek locally date from 2025 and mean ollama run deepseek-r1. That doesn't load V4, it loads an 8-billion-parameter model by default. Ollama lists these sizes for deepseek-r1:
DeepSeek calls these small models "distilled". They learned R1's reasoning pattern, but at their core they are Qwen or Llama models. That affects the license. According to Ollama, the Qwen variants derive from Qwen 2.5 under Apache 2.0, and the 70-billion-parameter variant from Llama 3.3, originally under its license. For DeepSeek-R1-0528-Qwen3-8B, the model card states MIT. If you want to use "DeepSeek locally" for client work, check the base model's license too. How different these licenses are is in Open Weight vs Open Source: Licenses.
Privacy: app, API or your own weights
I'm a developer, not a lawyer, and this isn't legal advice. What I compare are the providers' own statements, not a legal assessment of your case.
According to its privacy policy, DeepSeek has appointed a representative for the EU, and the policy lists rights such as access and erasure. That doesn't change the fact that with the app and API, your input goes to a country outside the EU. What that means for client data needs to be checked before you use it.
Run locally, DeepSeek is no different from any other open model in terms of data flow. Where your data goes isn't decided by where the model comes from, but by where it runs. Why "local" still doesn't automatically mean "privacy-compliant" is covered in Local AI and Privacy: Cloud or Server?.
When running DeepSeek locally pays off
Running DeepSeek V4 locally pays off if you already operate servers with several large GPUs or want to build a cluster of several machines. Then you get a large model under MIT without any data going to DeepSeek. For everyone else, two routes are more realistic. Either an open model between 20 and 35 billion parameters that runs on a single machine, such as Qwen3.8-27B, Muse Glimmer 30B or Gemma 4, all under Apache 2.0. Or a provider in the EU that runs open weights itself and offers a data processing agreement. The comparison of models for a single machine is in Open Source AI Models 2026 Compared, and my experience with the DGX Spark in DGX Spark Tested: Worth Buying?.
Frequently asked questions
Can I install DeepSeek V4 locally?
You may, the weights are on Hugging Face under MIT. In practice you need several GPUs or several linked machines: the V4-Flash files alone take up about 160 GB. On a single laptop, Mac or DGX Spark, V4 doesn't run in its published form.
How much memory does DeepSeek V4 Flash need?
The files in the repository take up about 160 GB, plus memory for the context. Nvidia lists models up to 200 billion parameters for a DGX Spark with 128 GB, and up to 400 billion for two linked Sparks. V4-Flash has 284 billion. I haven't tested it.
Is running DeepSeek locally GDPR-compliant?
Run locally, no input goes to DeepSeek, and that's the big difference from the app and API. Whether your whole setup is privacy-compliant depends on more than the model, though: on tools, logs and access. This isn't legal advice.
Can I use DeepSeek commercially?
The weights of V4 and V4.1 are under the MIT license, so commercial use is allowed. The condition is that you include the copyright and license notice when you pass the model on. For the small R1 models in Ollama, the base model's license also applies, for example Llama 3.3 for the 70B model.
What does "ollama run deepseek-r1" load?
By default DeepSeek-R1-0528-Qwen3-8B, an 8-billion-parameter model of 5.2 GB that builds on Qwen3 and was trained on R1's outputs. It's not DeepSeek V4 and not the full R1 with 671 billion parameters.
Where you go from here
If you're looking for an open model on your own hardware, start with the overview in Open Source AI Models 2026 Compared and choose by memory and license. I describe the setup with Ollama in Setting Up Local AI: Ollama to SGLang. And if you're thinking about buying hardware, do the math first with Your Own AI Server: Hardware and Cost.
