DeepSeek key facts
| What it is | A capable assistant available as a free hosted chat, a low-cost API, and open weights |
| Strongest at | Reasoning, mathematics and code — its models are explicitly trained toward these |
| Open weights | Yes — models can be downloaded and self-hosted, unlike the closed frontier labs |
| API pricing | Substantially below comparable Western APIs |
| Hosted service | Operated from China; data is processed under Chinese jurisdiction |
| Content handling | The hosted service applies content restrictions on politically sensitive topics |
| Self-hosted | Removes the jurisdiction question entirely — nothing leaves your infrastructure |
| Poor fit for | Regulated data on the hosted service, organisations with China-related restrictions |
Three different products, three different answers
The hosted chat
Free, capable, and processed on infrastructure in China. For general questions, learning, and code you would happily post publicly, that is unremarkable. For work material, client information or anything regulated, it is a decision that belongs with whoever owns data policy at your organisation, not with an individual user.
The API
The same jurisdictional question applies, at production volume rather than personal volume. The price difference is real and large enough to change what is economically viable to build — which is precisely why it deserves a deliberate decision rather than a default.
The open weights
This is the option that resolves most of the objections. Downloaded and run on your own hardware, or on a provider you choose, the model never contacts DeepSeek at all. No data leaves, no jurisdiction question arises, and the content restrictions applied by the hosted service are not in the picture.
The trade is the usual one for self-hosting: you need the hardware and the operational capability. Ollama, vLLM and LM Studio all run models of this family.
Deciding which one you actually want
| 1. Classify your data first | Public, internal, or confidential. This one question determines which of the three products is available to you, and it is not a technical question. |
| 2. Check organisational policy | Many enterprises and public bodies have explicit restrictions on Chinese-operated services. Find out before building, not after. |
| 3. Test reasoning tasks specifically | This is where the models are strongest. Evaluating them on creative writing measures the wrong thing. |
| 4. Probe the content restrictions | If your use touches history, politics or geography, test those queries directly on the hosted service before committing. |
| 5. Price the self-hosted path honestly | Free weights are not free to serve. GPU hire plus operational time frequently exceeds a cheap API for low volume. |
| 6. Do not mix the decisions | “DeepSeek is fine because it is open” is a common and wrong conflation — open weights say nothing about the hosted service you are actually using. |
DeepSeek against the alternatives
| What decides it | DeepSeek | Mistral Le Chat |
|---|---|---|
| Open weights | Yes | Yes, for several models |
| Hosted jurisdiction | China | France / EU |
| Cost position | Lowest | Competitive |
| Content restrictions | Present on hosted service | Fewer |
| Strongest at | Reasoning and code | Speed and European compliance |
| Choose when | Cost or self-hosting dominates | EU data residency is required |
Against the closed assistants — ChatGPT, Claude, Gemini — DeepSeek is competitive on reasoning and behind on ecosystem, tooling and polish. Very few people choose it for the chat interface. They choose it for the price, or for the weights.
Saying the awkward part plainly
The hosted service applies restrictions on topics the Chinese state treats as sensitive, and it processes data under Chinese jurisdiction. Both are verifiable, neither is a secret, and both are legitimate factors in a decision.
Equally: the technical achievement is real, the open-weights release is a genuine contribution that benefited everyone working in this field, and the cost reduction pressured the entire market in a direction users benefit from.
Both things are true. The sensible response is neither to dismiss the models nor to ignore where the hosted service runs — it is to match the deployment to the sensitivity of what you are doing.
Practical limitations
- Ecosystem is thin next to the established assistants — fewer integrations, less tooling.
- Hosted availability has been uneven, with capacity constraints during demand spikes.
- Content restrictions on the hosted service, which self-hosting removes.
- Documentation is lighter than Western equivalents.
- The usual confident wrongness applies as everywhere else.
Who DeepSeek is right for
Developers and researchers who want capable open weights they can actually run and modify. Teams whose API costs are the binding constraint on what they can build. Anyone doing reasoning-heavy or mathematical work on non-sensitive material. Self-hosters who want frontier-adjacent capability without a per-token bill.
Not for confidential or regulated data on the hosted service, for organisations with policy restrictions on Chinese services, or for anyone who needs a mature ecosystem around the assistant.
Other open-weight options
- Mistral Le Chat — European, open weights, EU data residency.
- Ollama — the simplest way to run these weights locally.
- vLLM — for serving them to many users at once.
- Open WebUI — a chat interface over models you host yourself.
Compiled from DeepSeek’s documentation and public sources. We have not hands-on tested this tool. Last reviewed 17 August 2026.
