---
title: "What is the seed parameter? Reproducible LLM outputs"
description: "The seed parameter asks an LLM API to sample the same way on repeat requests. How seed and system_fingerprint work, and why determinism is only best effort."
url: https://broadrouter.com/glossary/seed-parameter/
lang: en
updated: 2026-11-15
publisher: BroadRouter (a BroadNet site); API by BroadNet Inference, https://inference.broadnet.ai
---

# Seed parameter

Updated 2026-11-15

The seed parameter is an integer you send with a request so the API tries to sample the same tokens when the same request is repeated. OpenAI describes it as best effort: determinism is not guaranteed, and the system\_fingerprint field shows when the backend configuration has changed.

Also called: random seed, reproducible outputs, deterministic outputs, system\_fingerprint

Sampling is random on purpose. A seed fixes the randomness so a rerun has a chance of matching the first run.

## How it works

When [temperature](https://broadrouter.com/glossary/temperature/) is above zero, the server draws each next token at random from the model’s distribution. A random number generator drives those draws. Fix its starting value, the seed, and the same request should take the same draws.

OpenAI’s reference marks `seed` as Beta. The cookbook says “our system will make a best effort to sample deterministically”. To get matching outputs it says to keep these the same across requests:

-   the `seed` integer,
-   the prompt and messages,
-   `temperature`, [top\_p](https://broadrouter.com/glossary/top-p/) and `max_tokens`.

The response includes `system_fingerprint`, which OpenAI says “represents the backend configuration that the model runs with”. If it changes between two runs, expect different outputs even with the same seed.

```python
r = client.chat.completions.create(
    model=MODEL, seed=42, temperature=0.7,
    messages=[{"role": "user", "content": "Name a colour."}],
)
print(r.choices[0].message.content, r.system_fingerprint)
```

Open-model servers have the same knob. vLLM’s `SamplingParams` has `seed`, “Random seed to use for the generation”, with a default of none.

## Why it matters

-   **Evals.** When you compare prompts or models in an [eval](https://broadrouter.com/glossary/eval/), a fixed seed reduces noise between runs. Still run several samples; best effort is not a guarantee.
-   **Debugging.** A seeded request makes a strange reply easier to reproduce.
-   **Caching.** If you cache responses, include the seed in the cache key, as the [response caching post](https://broadrouter.com/blog/cache-llm-responses-cut-cost/) does.

## On BroadNet Inference

BroadNet’s [API reference](https://inference.broadnet.ai/docs?utm_source=broadrouter&utm_medium=referral&utm_campaign=glossary-seed-parameter&utm_content=body) does not document `seed` or `system_fingerprint`. Test it before relying on it: send the same seeded request five times and compare the outputs. If they differ, lower the temperature, store the outputs you need, and evaluate on several samples rather than one. The [migration walkthrough](https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/) lists seed among the OpenAI fields to test or remove.

## Common mistakes

-   Treating a seed as a promise of byte-identical output.
-   Changing max\_tokens or the system prompt between runs and blaming the seed for the difference.

## Questions

### Does the same seed always give the same answer?

No. OpenAI says "Determinism is not guaranteed" and that there is "a small chance that responses differ" even when the parameters and system\_fingerprint match.

### Is temperature 0 the same as setting a seed?

Not exactly. Low temperature makes the top token very likely; a seed fixes the random draws. Many teams use both, but neither alone guarantees identical output.

### Does BroadNet Inference support seed?

BroadNet's API reference does not document seed or system\_fingerprint. Send the same seeded request several times and compare before you rely on it.

## See also

-   [Temperature](https://broadrouter.com/glossary/temperature/) (term)
-   [Top-p (nucleus sampling)](https://broadrouter.com/glossary/top-p/) (term)
-   [Eval (LLM evaluation)](https://broadrouter.com/glossary/eval/) (term)
-   [Logprobs](https://broadrouter.com/glossary/logprobs/) (term)
-   [Chat Completions API](https://broadrouter.com/glossary/chat-completions-api/) (term)
-   [Frequency and presence penalty](https://broadrouter.com/glossary/frequency-and-presence-penalty/) (term)
-   [LLM benchmark](https://broadrouter.com/glossary/llm-benchmark/) (term)
-   [Cache LLM API responses to cut cost: an exact-match cache in Python](https://broadrouter.com/blog/cache-llm-responses-cut-cost/) (article)
-   [How to evaluate an LLM before switching providers](https://broadrouter.com/blog/evaluate-llm-before-switching/) (article)
-   [Migrating off the OpenAI API: a code walkthrough in Python, Node, cURL](https://broadrouter.com/blog/migrate-from-openai-api-code-walkthrough/) (article)

Try it on your real workload. New accounts start with $5 of credit for Gemma and transcription. DeepSeek, GLM and Qwen3.8 need a top-up.

[Get started](https://inference.broadnet.ai/account/?utm_source=broadrouter&utm_medium=referral&utm_campaign=glossary-seed-parameter&utm_content=page-cta) ·

**Sources**

-   [OpenAI API reference, create chat completion (seed, system\_fingerprint) (checked 2026-11-15)](https://developers.openai.com/api/reference/resources/chat/subresources/completions/methods/create)
-   [OpenAI Cookbook, reproducible outputs with the seed parameter (checked 2026-11-15)](https://developers.openai.com/cookbook/examples/reproducible_outputs_with_the_seed_parameter)
-   [vLLM documentation, SamplingParams (seed) (checked 2026-11-15)](https://docs.vllm.ai/en/latest/api/vllm/sampling_params/)
