Self-hosted Deepseek vs Deepseek API? What's the difference?

It's nearly two years since DeepSeek rocked Wall Street, releasing OpenWeight models that challenged Frontier labs. Since then, we've seen an influx of Chinese model providers at a fraction of the cost, and the emergence of US and European OpenWeight models coming back into the fight.

Is the model good enough for the job? Where does the inference take place? The first has a short answer. The second is what this post is about.

PrivateMind will be serving DeepSeek-V4.1-Flash-1017 next week.

The Capability question is mostly settled

DeepSeek-V4.1-Flash, released in September 2026 (further updates with 1017 released). It's MIT-licensed, the weights are on Hugging Face, and scores 90.9 on GPQA Diamond. On agentic work it sits at the top of the published tables: 74.2 on DeepSWE and 31.8 on Agent's Last Exam, ahead of the leading closed models on both. The reason firms keep asking about DeepSeek is that the models keep earning the question.

DeepSeek API

When you use a provider's application or API such as DeepSeek's own, your prompts, files and chat history are sent to DeepSeek and stored in China. DeepSeek's privacy policy for reference: https://cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html.

None of this is hidden, and similar providers worldwide have identical policies in place. The difference is the jurisdiction. Data is held in China, under different laws that US and EU firms have no practical way to audit or challenge. Data can be processed and trained on to improve models unless you opt out.

For regulated firms, that is the end of the conversation about the hosted service. Client data, IP and anything sensitive should not be sent to any provider without knowing the implications.

What changes when you host the weights?

The model weights from DeepSeek are publicly available, from sites such as HuggingFace (recently acquired by Nvidia for $12.9 Billion). The model is downloaded once, under a permissive license and runs on your infrastructure. Prompts and data are processed on the GPUs you control. Only you see the data processing and have control of where that data is sent to.

	Public API providers	Self-hosted weights
Where does your data go?	Sent to the provider, often overseas	Stays on your servers
Who controls your data?	The provider	You do
What happens to it?	Kept per their policy	Kept per your policy
When does it change?	When they update it	When you update it

Running the weights yourself allows you to control the parameters and run without any outbound access, ensuring data never leaves your network.

In short, self-hosting turns the data sensitivity question into an ordinary infrastructure question. The ability to run larger 100B+ parameter models becomes a HBM issue, both in capacity and supply chain constraints.

Guardrails are someone else's decision

Another elephant in the room around is the use of different models and the guardrails. Take the Hugging Face incident in July 2026 where an autonomous evaluation agent escaped its sandbox and ran a days-long intrusion against Hugging Face's production infrastructure. 

Investigators tried to use closed frontier models to analyse the attack; the models refused. Their guardrails treated reverse-engineering an exploit the same as launching one. The team stood up an open-weight model, GLM-5.2, on its own infrastructure and ran the entire forensic pipeline through it instead. Someone else's guardrails are a dependency you cannot negotiate away, and legitimate work gets blocked by them at the worst possible moment.

Where does this leave regulated financial services?

The model and the service are two separate decisions. Is DeepSeek safe to use? On your own infrastructure, yes. As a hosted API, that depends entirely on where the data lands. Can we run it ourselves? Yes, and that is the version of the question worth answering but depends on the capacity of infrastructure you have available.

What does this mean for PrivateMind?

Sovereign AI and privacy were the founding premises of PrivateMind. The platform can run open-weight models like DeepSeek on pooled or dedicated infrastructure inside your jurisdiction: weights pinned to the revision you approve, no egress unless you allow it, and every prompt landing on GPUs with logs you control. The question was never whether you can use DeepSeek. It is who is holding the data when you do.

Check out privatemind.com and let us know your thoughts.