Try DeepSeek V4.1 Flash 1017 on three practical tasks: reviewing code, extracting JSON, and comparing documents. Each example includes a prompt and a short checklist for reviewing the answer. Copy it into the playground, or send it through the API.
Review a function
Ask for one reproducible problem before requesting a broad rewrite. A small task makes it easier to distinguish a correct fix from unrelated changes.
Review this Python function. Explain what happens for an empty list,
show one failing input, and propose the smallest fix. Include two tests.
State any assumptions about the intended behavior.
Do not silently decide what an empty list should return. Explain the
choice you make or ask for clarification.
def average(values):
return sum(values) / len(values)
What to check
The answer should identify the division-by-zero problem for an empty list. It should distinguish finding the bug from deciding the intended empty-input behavior. Check that the proposed tests cover both a normal input and the chosen empty-input rule.
Run the suggested code and tests. A test that merely repeats the implementation is weaker evidence than one that checks an independently specified result.
A useful follow-up
Assume empty input must raise ValueError with a clear message.
Update only what is necessary and add a test for that requirement.
Explain whether the existing normal-input behavior changes.
This follow-up supplies a concrete requirement instead of asking the model to guess the product decision.
Extract structured data
Use a short source with deliberately missing information. This tests whether the response follows the requested format without inventing details.
Extract tasks from these notes. Return only JSON with a tasks array.
Each task must have title, owner, and due_date.
Use null when a field is not stated. Do not infer a date or owner.
Dates, when provided, must use YYYY-MM-DD.
Notes:
Maya will update the API guide.
Review the onboarding copy by November 5, 2026; no owner has been assigned.
What to check
The result should contain two tasks. The API-guide task has Maya as its owner and no stated due date. The onboarding-review task has a stated date of 2026-11-05 and no assigned owner.
Validate that the output is parseable JSON and that missing fields use null. Check the field values against the notes. Asking for JSON in a prompt does not itself establish that an endpoint enforces a schema.
A useful follow-up
Check the extracted tasks against the original notes.
List any value that is not explicitly supported by the notes.
Do not fill missing values.
For application use, validate JSON in your application as well. A second model response cannot replace a parser or field checks.
Compare document sections
Ask the model to identify a contradiction and point to the text that supports it. Keep decisions about the final policy outside the task.
Compare these two sections. Identify contradictions and quote the
relevant sentence from each section. List what needs clarification.
Do not choose a policy on our behalf.
Section A: Trial credits expire after seven days.
Section B: Trial credits never expire.
The text above is a synthetic exercise, not this service's trial policy.
What to check
The response should identify the conflicting expiry rules, quote both statements accurately and explain that a policy decision is required. It should not invent an exception that reconciles the sections or select a rule without being asked.
For a longer document, provide stable section names and check every quotation against the original. If the source does not answer a question, the answer should preserve that uncertainty.
Coding tools and agent workflows
A model, a chat interface and a coding agent are different parts of a workflow. The model produces responses; the surrounding software determines which files, commands and tools it can use. That context matters when comparing demonstrations.
Users discussing 1017 have asked which interface produced a game, requested the original prompt and questioned whether a Web UI comparison used comparable conditions. These are useful questions to ask of any demonstration. Community comparison.
When evaluating OpenCode, DeepSeek Harness or another coding tool:
- Record the tool version, provider and selected model.
- Check that the endpoint supports the request format and features the tool uses.
- Use a small test project with clear success criteria.
- Review file changes and test output, not just the assistant's final message.
- Record retries, context compression and total usage.
Our API accepts text chat without tool calls or streaming. Coding agents that require those features cannot use this endpoint as a drop-in replacement. Check the supported request format before connecting a tool.
Build your own comparison
Choose a few tasks that resemble your actual work. Keep inputs and acceptance criteria fixed, record configuration changes and retain both successful and failed attempts. A visual demonstration can help you discover a capability, while a reproducible test helps you decide whether to depend on it.
Use a simple record with these fields:
| Field | What to record |
|---|---|
| Task | The input and expected outcome |
| Configuration | Model, provider, date, settings and tools |
| Result | Passed checks, failed checks and manual corrections |
| Time | Time to initial output and total completion time |
| Usage | All requests and available billing records |
| Decision | Whether the result is useful for this workflow |
Read the API guide · Understand task costs · Open the playground