Skip to content
Multigrid

DeepSeek model behaviour

DeepSeek's reasoning models and their distilled variants: how reasoning content is exposed, why the system prompt behaves unusually, and what each size actually holds.

DeepSeek’s API is deliberately OpenAI-shaped: the same base path, the same messages array, the same choices[0].delta in a stream. That similarity is useful and it is also where the surprises come from, because the places DeepSeek differs are not announced by a different-looking response. A reasoning model returns an extra field alongside content. Four sampling parameters are accepted and discarded while two others raise a 400. A caching layer you never asked for changes the bill and reports itself in two usage counters that have no equivalent elsewhere.

The second thing to hold on to is that “DeepSeek-R1” names two different things — a hosted endpoint and a set of open weights — and the reasoning trace is exposed differently by each. These pages work through the observable behaviour of both: what is in the response, which parameters do what, what the model card and the licence actually say, and where the documented value is expected to move.

DeepSeek-V3's Context Window and Output Limit

The context length DeepSeek publishes for V3, where the number is recorded, and why the output ceiling is a separate and much smaller figure.

8 min read

DeepSeek-R1's Reasoning Tokens and How They're Billed

Where reasoning tokens appear in the usage object, why they are charged at the output rate, and a worked bill separating them from the answer you keep.

8 min read

Parsing DeepSeek-R1's think Block Out of the Response

A parser that separates R1's reasoning trace from its answer, handling both the hosted API's separate field and the open weights' literal tags.

9 min read

Function Calling in DeepSeek: Which Models and What Format

The tool-call request and response shape DeepSeek documents, which endpoints expose it, and the instability DeepSeek warns about in its own docs.

8 min read

Why DeepSeek-R1 Ignores the System Prompt

Your system message is accepted without error and has no visible effect on R1; what DeepSeek actually recommends, why the reasoning phase swamps it, and where to put the instruction instead.

8 min read

DeepSeek's Tokenizer and Vocabulary Size

The vocabulary size DeepSeek ships, the special tokens that make its templates unusual, and how to count tokens exactly rather than estimating.

8 min read

Context Caching on Disk in the DeepSeek API

DeepSeek caches prompt prefixes on disk automatically, reports hits and misses in two usage fields, and discounts the hit portion of your input.

8 min read

The Streaming Response Shape of the DeepSeek API

The SSE envelope DeepSeek emits, what a delta contains, how reasoning content arrives on its own key, and the finish_reason value that has no OpenAI equivalent.

8 min read

DeepSeek's Model Licence: What the Terms Permit

Which licence applies to DeepSeek's open-weight releases, why code and weights are licensed separately, and why the distilled variants are governed by someone else's terms.

8 min read

DeepSeek-V3's Mixture-of-Experts Architecture and Active Parameters

The total and active parameter counts DeepSeek publishes for V3, what 'active' means for a single token, and which of the two numbers predicts cost.

8 min read

Stop Sequences in the DeepSeek API

The stop parameter's shape and limit, why the matched sequence is removed from the output, and why finish_reason cannot tell you a stop sequence fired.

7 min read

DeepSeek's max_tokens Default and Output Ceiling

The default and maximum output length DeepSeek documents for each endpoint, why the reasoning endpoint's are far larger, and what truncation looks like.

8 min read

DeepSeek's Documented Content Policy Restrictions

Where DeepSeek's rules for the hosted API are published, what signal a blocked request returns, and why self-hosted weights are governed by something else entirely.

8 min read

JSON Mode Output in the DeepSeek API

How to enable DeepSeek's JSON output mode, the two conditions its documentation attaches to it, and the failure modes it does not protect you from.

8 min read

DeepSeek-R1's Distilled Variants: Context Window by Size

Each R1-distill inherits its architecture and its context window from a Qwen or Llama base, which is why the sizes do not share one number.

9 min read

Why DeepSeek-R1's Reasoning Trace Varies Run to Run

The same prompt produces a different think block every time because the reasoning phase is sampled and you cannot turn the sampling off from the API.

8 min read

DeepSeek's Two Model Names: deepseek-chat and deepseek-reasoner

Two model names on one base URL that differ in accepted parameters, response fields, output ceilings and feature support — and what changes in your client for each.

8 min read

Why DeepSeek-R1 Ignores the Temperature Parameter

Temperature is accepted without error on the reasoning endpoint and has no effect; which parameters error instead, why the sampling is fixed, and what you can still control.

8 min read

Other topics