In DeepSeek's API, what do the deepseek-chat and deepseek-reasoner names select?

## The two names DeepSeek's hosted API does not present a long catalogue of dated model IDs the way some vendors do. It presents two working names: - **`deepseek-chat`** — the general-purpose line, descended from the DeepSeek-V3 series. It answers directly: you send messages, you get an assistant message back. Use it for ordinary conversation, summarisation, extraction, classification, code writing, and anything where you want the answer without a deliberation phase. - **`deepseek-reasoner`** — the reasoning line, descended from DeepSeek-R1. Before producing the user-visible answer it generates a chain of thought, and that thinking is itself generated text. It is the choice for problems where the model genuinely has to work something out: competition-style maths, formal or semi-formal proofs, multi-constraint planning, tricky root-cause debugging. Both lines come from the same open-weight programme. DeepSeek publishes the underlying weights, so the hosted names and the downloadable checkpoints are two views of one family rather than two unrelated products. ## Alias, not a version pin The single most important property of these names for anyone running production traffic is that they are **aliases**. `deepseek-chat` means "the current general chat model", not "the exact weights that shipped on some date". When DeepSeek releases a new generation of the V3 line, the alias moves. Nothing in your code changes; the model behind it does. That is convenient — you get improvements for free, without a migration — and it is a hazard, because prompt behaviour, output formatting, refusal boundaries and even token counts can move underneath a system you have already tuned. Treat the alias as a floating dependency, the same way you would treat a `latest` container tag. ## What the names are not - They are **not** model architectures. Sparse-expert structure, parameter counts and attention design belong to the checkpoint, not to the name you type. - They are **not** the Hugging Face repository names. `deepseek-ai/DeepSeek-R1` is a weight repository; `deepseek-reasoner` is an API alias. Sending the repository name as the `model` value is a common first-day mistake and simply fails. - They are **not** tiers of one model in the way a vendor's small/medium/large ladder is. Choosing between them is choosing a *behaviour* — deliberate versus direct — not a size on a price ladder. ## Choosing between them in practice Start with `deepseek-chat`. It is faster and it does not spend output tokens on deliberation. Move a specific call path to `deepseek-reasoner` only when you can show, on your own evaluation set, that the extra thinking actually improves the result for that path. "Use the smart one everywhere" is a real anti-pattern: it inflates latency and output-token spend on tasks such as JSON reformatting or short classification where deliberation buys nothing. Because the two names accept the same request shape, switching a call path is usually a one-line change, which makes A/B evaluation on your own data cheap. Take advantage of that instead of arguing about it in design review. ## Pinning If you need behaviour to be stable — a regulated workload, a benchmark you must reproduce, a contract that promises consistent output — the hosted alias cannot give you that guarantee. The escape hatch is specific to DeepSeek's open-weight strategy: download a named revision of the weight repository and serve it yourself, or use a host that lets you address one exact checkpoint. You then own the upgrade decision, at the cost of owning the serving stack. ## Version note As of mid-2026, `deepseek-chat` and `deepseek-reasoner` are the two primary names on DeepSeek's own API. Because the aliases float, quote the *behaviour* of each line in an interview rather than reciting a point-version number, which will be stale within months.