What is FIM
FIM (Fill-in-the-Middle) is a special completion mode. Traditional text generation can only continue from left to right, while FIM allows the model to fill in the missing "middle" part based on the "preceding context" and "following context". This is extremely useful in code completion scenarios—you write the first half and the second half of a function signature, and the model helps you complete the implementation in between.
FIM Completion API
Both DeepSeek V4 Flash and Pro support FIM completion (non-thinking mode only):
from openai import OpenAI
import os
client = OpenAI(
api_key=os.environ.get('DEEPSEEK_API_KEY'),
base_url='https://api.deepseek.com'
)
response = client.completions.create(
model='deepseek-v4-flash',
prompt='def fibonacci(n):\n if n <= 1:\n return n',
suffix='\n\n# テスト\nprint(fibonacci(10))',
max_tokens=256,
temperature=0,
stop=['\n\n#']
)
print(response.choices[0].text)
# 出力：
# return fibonacci(n-1) + fibonacci(n-2)Prompt Design Best Practices
| Strategy | Example | Effect |
|---|---|---|
| Function signature hint | Preceding context = function signature + docstring, following context = return statement | Complete the function body implementation |
| Test-driven | Preceding context = function definition, following context = test cases | Generate code that satisfies the tests |
| Comment-driven | Preceding context = comment description, following context = type annotations | Generate code according to comments |
Integration into Editors
VS Code (via Continue plugin): Configure tabAutocompleteModel to point to DeepSeek to enable FIM completion.
Neovim (via llm.nvim): Configure the fim endpoint to point to the DeepSeek completions API.
FIM vs Chat API Comparison
| FIM Completion | Chat Completion | |
|---|---|---|
| Endpoint | /v1/completions | /v1/chat/completions |
| Context | Preceding + following context | Only preceding context (conversation history) |
| Speed | Faster (no conversation overhead) | Slower |
| Cost | Lower (no system messages) | Higher |
| Use cases | Real-time completion, inline suggestions | Conversational programming, complex Q&A |
Notes
- FIM only supports non-thinking mode; enabling thinking will result in an error.
- The suffix parameter is required (unlike the standard completions API).
- It is recommended to set the stop parameter to the first few characters of the following context to avoid over-generation.
- This is currently a Beta feature; the API may be subject to changes.