Updated 2026-06-29
DeepSeek FIM Completion (Beta): use prefix and suffix deliberately, not like ordinary chat
DeepSeek's official FIM Completion (Beta) page covers a different workflow from chat completions. Instead of asking for a whole answer in conversation format, you send a prefix, optionally send a suffix, and ask the model to fill the gap in between. The page also adds two practical constraints many quick tutorials miss: the feature runs on the Beta base URL, and DeepSeek caps FIM completion output at 4K tokens. That makes this a strong long-tail page for editor-completion and code-insertion intent.
1. FIM means the model completes the middle between what you already know
DeepSeek describes FIM as Fill In the Middle completion. The pattern is useful when you already know the start of the content and possibly the end, but want the model to generate the missing section.
That is why the official sample uses a function prefix and a recursive-return suffix. The model is not inventing the whole answer from scratch; it is fitting code into an explicit structural gap.
response = client.completions.create(
model="deepseek-v4-pro",
prompt="def fib(a):",
suffix=" return fib(a-1) + fib(a-2)",
max_tokens=128,
)Sources checked
- DeepSeek official FIM Completion (Beta) guide - Primary source for the prefix-and-suffix flow, 4K cap, beta base URL, and Continue mention.
2. The feature is Beta-only and uses the completions API shape
DeepSeek's official sample calls `client.completions.create`, not the chat-completions path. That matters because FIM is framed as a completion workflow rather than a conversational turn.
The docs also require `base_url=https://api.deepseek.com/beta` to enable the feature. If you send the right prompt and suffix to the stable route, you are outside the documented Beta contract.
client = OpenAI(
api_key="<your api key>",
base_url="https://api.deepseek.com/beta",
)3. The official output ceiling is 4K tokens
DeepSeek's Notice section sets a hard practical boundary: the maximum tokens for FIM completion is 4K.
That means FIM is better suited to focused code insertion, template filling, or bounded content completion than to huge middle-of-document rewrites. If your gap is massive, you need to redesign the prompt or split the task.
| Check | What to verify | Why it matters |
|---|---|---|
| 1 | Use the Beta base URL | FIM is documented as a Beta feature |
| 2 | Send a clear prefix and optional suffix | The model needs an explicit gap to fill |
| 3 | Keep `max_tokens` within the 4K limit | The official docs cap FIM completion output at 4K |
| 4 | Use a bounded insertion task | FIM is stronger for code or content gaps than for broad rewrites |
4. Continue is the editor-integration clue, not the whole product story
DeepSeek's page explicitly mentions Continue as a VS Code plugin that supports code completion and points readers to Continue-specific configuration for this feature.
That is useful product intent, but the DeepSeek-first story stays on the API capability itself: the official contract is still Beta base URL plus prefix/suffix completion logic, regardless of which editor wrapper calls it.
5. Keep FIM separate from chat and tool-calling questions
If your real problem is inserting content into a partially known code structure, FIM is the right official topic. If your problem is multi-step reasoning with external tools, compare `/guides/deepseek-thinking-mode-tool-calls` instead.
If the task is controlling the opening of an assistant reply in chat rather than filling a middle gap, continue with `/guides/deepseek-chat-prefix-completion-beta`.
6. Do not confuse Beta completion features with stocked products
FIM Completion is an API feature, not a separate product listing. It does not authorize any `/pricing` inventory change or new plan card.
Use `/pricing` only when the user needs an in-stock DeepSeek key route that the site actually sells today.
FAQ
What is DeepSeek FIM Completion?
It is DeepSeek's Fill In the Middle Beta feature, where you provide a prefix and optionally a suffix, and the model completes the content in between.
Does DeepSeek FIM use the normal stable base URL?
No. The official docs require `https://api.deepseek.com/beta` to enable the feature.
What is the maximum token budget for DeepSeek FIM completion?
The official FIM guide says the max tokens for FIM completion is 4K.
Is FIM better for code insertion than ordinary chat?
Yes, when you already know the prefix and possibly the suffix. The official design is built around filling a bounded gap.
Does this page mean DeepSeek sells a separate FIM plan on `/pricing`?
No. This page documents an API feature only. Purchasable plans still depend on actual stocked inventory.
The practical DeepSeek FIM rule is simple: use the Beta route, send a clear prefix and optional suffix, keep the completion inside the 4K cap, and reserve the pattern for bounded insertion tasks where filling the middle is the real job.
Related model comparisons
Continue from this guide into structured DeepSeek-first comparison pages with model tables, routing advice, and pricing context.