DeepSeek is a Chinese AI company whose models became widely known for one reason: they perform close to the frontier while being dramatically cheaper, and the weights are openly available.
That combination is unusual, and it is why DeepSeek comes up in nearly every conversation about AI cost.
What makes it different
Open weights. You can download the model and run it on your own hardware. No closed provider allows this. It means no per-token cost, no data leaving your infrastructure, and no vendor able to change terms on you.
Cost. Where DeepSeek is offered as a hosted API, pricing is typically an order of magnitude below the closed frontier models. For high-volume work this changes what is economically possible, not just what is cheaper.
Reasoning. Its reasoning-focused models show their working — you can watch the chain of thought rather than just receiving a conclusion. Useful for checking, and genuinely interesting to read.
Competitive quality. Not top of every benchmark, but close enough that for many real tasks the difference does not show.
What it is good at
- Reasoning and mathematics — a consistent strength
- Code — well regarded, particularly given the price
- High-volume work where per-token cost dominates
- Self-hosted deployments where data cannot leave your infrastructure
Where it falls short
Product polish. The experience around the model is rougher than the closed alternatives. Fewer features, less integration, less consistency between the services hosting it.
Writing. Capable but not the strongest. For prose a person reads closely, Claude remains the better tool.
Consistency across hosts. Because the weights are open, "DeepSeek" through one provider may behave differently from another — different quantisation, different settings, different speeds. Your experience depends on who is running it.
Censorship on the official service. The hosted Chinese service declines politically sensitive topics, particularly regarding China. Self-hosted or Western-hosted deployments behave differently. Worth knowing which you are using before drawing conclusions about the model.
The data question
Reasonable and worth answering plainly.
Using the official DeepSeek service means sending your data to servers operating under Chinese jurisdiction. For many organisations that is a straightforward no, regardless of the model's quality.
Using DeepSeek through a Western host, or running it yourself, does not involve that at all — the weights are just files. This distinction gets lost in most discussion, and it matters: the concern is about the service, not the model.
If your objection is jurisdictional, a different host resolves it. If your objection is to the model's training or its refusal behaviour, it does not.
Open weights, not open source
A small precision that gets muddled constantly.
DeepSeek releases weights — the trained parameters. That lets you run and fine-tune the model. It does not release the full training data and pipeline, which is what "open source" would normally imply in software.
This is the same position as most so-called open models, including Llama. "Open weights" is the accurate term. See open-source AI models for what that means in practice.
Should you use it?
Yes, if cost is your binding constraint, you need to self-host for data reasons, or you are doing high-volume reasoning or code work.
Probably not, if you want the most polished experience, the best prose, or the fewest moving parts — and you are only running a modest volume where the cost difference is a rounding error.
Most people: worth having access to alongside the closed models rather than instead of them. It is genuinely better at some things and genuinely rougher at others.
Common questions
What is DeepSeek? A Chinese AI company whose open-weight models perform near the frontier at dramatically lower cost.
Is DeepSeek free? The weights are freely available to download and run. Hosted access is usually free or very cheap depending on the provider.
Is DeepSeek safe to use? Depends what you mean. The official service sends data to Chinese-jurisdiction servers. Self-hosted or Western-hosted deployments do not.
Is DeepSeek better than ChatGPT? Competitive on reasoning and code at far lower cost. Behind on product polish and prose quality.
Is DeepSeek open source? It releases open weights, not full training data and pipeline. "Open weights" is the accurate description.
Can I run DeepSeek on my own computer? Smaller variants, yes, with adequate hardware. The largest models need serious infrastructure.
Why is it so much cheaper? A combination of efficient architecture and a deliberate pricing position. Open weights also mean many providers compete to host it, which pushes prices down.