What do temperature, top_p and seed do?

Temperature and top_p control how a model picks each next token: lower values make output focused and repeatable, higher values make it more varied. A seed fixes the random choices, so the same request with the same seed returns the same or a very similar result.

Concepts

How a model picks the next token

A language model writes one token at a time. At each step it assigns a probability to every token in its vocabulary, and a sampling step picks one of them. Temperature and top_p reshape those probabilities before the pick, and a seed fixes the random numbers the pick uses. Together they decide how predictable or varied the output is.

Temperature

Temperature scales how strongly the model prefers its most likely tokens. Low values concentrate the choice on the top candidates, which makes output focused and consistent. High values spread the choice across more candidates, which makes output more varied and, past a certain point, less coherent. At 0, the model almost always takes the single most likely token.

  • 0 to 0.3: extraction, classification, code and factual answers.
  • 0.5 to 1.0: general conversation and writing.
  • Above 1.0: brainstorming and varied drafts, with closer review of the results.

Accepted ranges differ by model: some take 0 to 1 and many go up to 2. When you leave temperature out, the model's own default applies. On some speech models, temperature sets how expressive and variable the voice sounds rather than which words appear. Some reasoning models use fixed sampling settings and ignore temperature, and their model pages say so.

top_p

top_p, also called nucleus sampling, limits the choice to the smallest set of tokens whose probabilities add up to the value you set. At 1, no token is excluded. At 0.9, the least likely tokens that together hold the last 10% of probability are removed before sampling. Lower values focus the output further.

Adjust temperature or top_p, usually not both, since changing both at once makes the effect hard to predict. Some models also accept top_k, which keeps only the K most likely tokens.

Seed

Sampling relies on random numbers. A seed fixes their starting point, so sending the same request with the same seed and settings tends to repeat the same choices. That helps when you test prompts, debug an application or compare two versions of a prompt fairly.

A seed is not a guarantee on text models. Model pages for many text models describe the result as similar, not always identical: small numerical differences between runs, a newer model version, or any change to the input or settings can alter the output. On some models a seed is not applied when web search is on. When you need exactly the same text again, store the response instead of generating it again.

Many image, video and audio generation models accept a seed too. Reusing it with the same prompt and settings returns the same result or a close variation, depending on the model, which makes it easier to refine one prompt step by step.

Testing settings

Change one setting at a time and compare several outputs for the same prompt, since a single sample can mislead. For tasks with one correct answer, start low and raise temperature only if the output becomes repetitive. For open-ended writing, start near the middle of the model's range and adjust from there. The models listed on this page show which of these parameters each one accepts. Try values in the Playground before you set them in code, and read about reasoning models, some of which fix their sampling settings.

Models that accept temperature, top_p, seed

113 models. Each model's page has its rate card, parameters and code.

ModelParameterRange / values
Gemma 4 26B-A4B Googletemperature top_p seedtemperature: 0 to 2; default 1
top_p: 0.01 to 1; default 0.95
seed: 0 to 2147483647
Muse Spark 1.1 Meta AItemperature top_p seedtemperature: 0 to 2; default 1
top_p: 0.01 to 1; default 1
seed: -
Muse Spark 1.2 Meta AItemperature top_p seedtemperature: 0 to 2; default 1
top_p: 0.01 to 1; default 1
seed: -
Qwen3.5 122B-A10B Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.5 27B Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.5 35B-A3B Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.5 397B-A17B Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.5 4B Alibaba Cloudtemperature top_p seedtemperature: 0 to 2; default 0.7
top_p: 0.01 to 1; default 0.95
seed: 0 to 2147483647
Show all 113 models
ModelParameterRange / values
Qwen3.5 9B Alibaba Cloudtemperature top_p seedtemperature: 0 to 2; default 0.7
top_p: 0.01 to 1; default 0.95
seed: 0 to 2147483647
Qwen3.5 Flash Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.5 Omni Flash Realtime Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99
top_p: 0.01 to 1
seed: 0 to 2147483647
Qwen3.5 Omni Plus Realtime Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99
top_p: 0.01 to 1
seed: 0 to 2147483647
Qwen3.5 Plus Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.6 27B Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.6 35B A3B Alibaba Cloudtemperature top_p seedtemperature: min 0; default 0.7
top_p: 0.01 to 1; default 0.95
seed: 0 to 2147483647
Qwen3.6 Flash Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.6 Plus Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.7 Flash Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.7 Max Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.7 Plus Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.8 27B Alibaba Cloudtemperature top_p seedtemperature: 0 to 2; default 1
top_p: 0.01 to 1; default 0.95
seed: 0 to 2147483647
Qwen3.8 Flash Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.8 Max Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.8 Max 0902 Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.8 Omni Flash Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
Qwen3.8 Omni Flash Realtime Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99
top_p: 0.01 to 1
seed: 0 to 2147483647
Qwen4 Max Alibaba Cloudtemperature top_p seedtemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
seed: 0 to 2147483647
SoulX Podcast Soul AI Labtemperature top_p seedtemperature: 0.1 to 2; default 0.6
top_p: 0.1 to 1; default 0.9
seed: default 42
DeepReasoning WinFunctemperature top_ptemperature: 0 to 2; default 0.7
top_p: 0 to 1; default 0.9
DeepSeek V4 Flash 0731 DeepSeektemperature top_ptemperature: 0 to 2; default 0.7
top_p: 0 to 1; default 0.9
DeepSeek V4 Pro DeepSeektemperature top_ptemperature: 0 to 2; default 0.7
top_p: 0 to 1; default 0.9
DeepSeek V4 Pro 0813 DeepSeektemperature top_ptemperature: 0 to 2; default 0.7
top_p: 0 to 1; default 0.9
DeepSeek V4.1 Flash 1017 DeepSeektemperature top_ptemperature: 0 to 2; default 0.7
top_p: 0 to 1; default 0.9
GLM 4.5 Flash Z.aitemperature top_ptemperature: 0 to 1; default 0.6
top_p: 0.01 to 1; default 0.95
GLM 4.6V Flash Z.aitemperature top_ptemperature: 0 to 1; default 1
top_p: 0.01 to 1; default 0.95
GLM 4.6V FlashX Z.aitemperature top_ptemperature: 0 to 1; default 1
top_p: 0.01 to 1; default 0.95
GLM 4.7 Flash Z.aitemperature top_ptemperature: 0 to 1; default 1
top_p: 0.01 to 1; default 0.95
GLM 4.7 FlashX Z.aitemperature top_ptemperature: 0 to 1; default 1
top_p: 0.01 to 1; default 0.95
GLM 5.1 Z.aitemperature top_ptemperature: 0 to 1.99; default 1
top_p: 0 to 1; default 0.95
GLM 5.2 Z.aitemperature top_ptemperature: 0 to 1; default 1
top_p: 0.01 to 1; default 0.95
GLM 5.3 Z.aitemperature top_ptemperature: 0 to 1; default 1
top_p: 0.01 to 1; default 0.95
GLM 5.3 Flash Z.aitemperature top_ptemperature: 0 to 1; default 1
top_p: 0.01 to 1; default 0.95
Gemma 3 27B Googletemperature top_ptemperature: 0 to 2; default 1
top_p: 0.01 to 1; default 1
Kimi K2.6 Moonshot AItemperature top_ptemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
MiMo V2.6 Flash Xiaomitemperature top_ptemperature: 0 to 1.5; default 1
top_p: 0.01 to 1; default 0.95
MiMo V2.6 Pro Xiaomitemperature top_ptemperature: 0 to 1.5; default 1
top_p: 0.01 to 1; default 0.95
MiMo V2.6 Pro UltraSpeed Xiaomitemperature top_ptemperature: 0 to 1.5; default 1
top_p: 0.01 to 1; default 0.95
Mistral Small 3.1 Mistral AItemperature top_ptemperature: 0 to 5; default 0.15
top_p: 0.01 to 1; default 1
Muse Glimmer 30B Meta AItemperature top_ptemperature: min 0; default 1
top_p: 0.01 to 1; default 0.95
Muse Spark 1.3 Meta AItemperature top_ptemperature: 0 to 2; default 1
top_p: 0.01 to 1; default 1
Nova Lite 1.0 Amazontemperature top_ptemperature: 0 to 1; default 0.7
top_p: 0 to 1; default 0.9
Nova Lite 2 Amazontemperature top_ptemperature: 0 to 1; default 0.7
top_p: 0 to 1; default 0.9
Nova Micro 1.0 Amazontemperature top_ptemperature: 0 to 1; default 0.7
top_p: 0 to 1; default 0.9
Nova Pro 1.0 Amazontemperature top_ptemperature: 0 to 1; default 0.7
top_p: 0 to 1; default 0.9
Qwen3.5 Omni Flash Alibaba Cloudtemperature top_ptemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
Qwen3.5 Omni Plus Alibaba Cloudtemperature top_ptemperature: 0 to 1.99; default 0.7
top_p: 0 to 1; default 0.9
Seed 2.0 Mini ByteDancetemperature top_ptemperature: 0 to 2; default 0.7
top_p: 0 to 1; default 0.9
Step 3.5 Flash StepFuntemperature top_ptemperature: 0 to 2; default 0.5
top_p: 0.01 to 1; default 0.9
Step 3.5 Flash 2603 StepFuntemperature top_ptemperature: 0 to 2; default 0.5
top_p: 0.01 to 1; default 0.9
Step 3.7 Flash StepFuntemperature top_ptemperature: 0 to 2; default 0.5
top_p: 0.01 to 1; default 0.9
Step 5 Preview StepFuntemperature top_ptemperature: 0 to 2; default 0.5
top_p: 0.01 to 1; default 0.9
StepAudio 2.5 Chat StepFuntemperature top_ptemperature: 0 to 2; default 0.5
top_p: 0.01 to 1; default 0.9
StepAudio 3 Chat StepFuntemperature top_ptemperature: 0 to 2; default 0.5
top_p: 0.01 to 1; default 0.9
StepAudio 3 Music StepFuntemperature top_ptemperature: 0.01 to 2; default 0.85
top_p: 0 to 1; default 0.92
ACE-Step 1.5 XL ACE-Stepseed-
FLUX.2 Klein 4B Black Forest Labsseed-
GLM TTS Z.aiseed-
Gemini 3.8 Live Googletemperature0.2 to 2
Gemini 3.8 Live Extended Thinking Googletemperature0.2 to 2
HappyHorse 1.0 Alibaba Cloudseed-
HappyHorse 1.1 Alibaba Cloudseed0 to 2147483647
Hunyuan Image 3 Tencentseed-
Hunyuan Video 1.5 Tencentseedmin 0
Hyper3D Gen2 Deemosseed0 to 65535
MOSS Video and Audio OpenMOSSseed0 to 4294967295
MOSS-Image Turbo OpenMOSSseed0 to 2147483647
MiniMax Music 3 MiniMaxseed-
Mistral Small 4 Mistral AItemperature0 to 1; default 0.7
OpenAI Whisper 1 OpenAItemperature0 to 1; default 0
PixVerse V6 Max PixVerseseed0 to 2147483647
Pixverse C1 PixVerseseed0 to 2147483647
Pixverse Edit PixVerseseed0 to 2147483647
Pixverse v5 PixVerseseed-
Pixverse v5.6 PixVerseseed-
Pixverse v6 PixVerseseed0 to 2147483647
Qwen Audio 3.0 TTS Alibaba Cloudseed0 to 65535; default 0
Qwen Audio 3.1 TTS Alibaba Cloudseed0 to 65535; default 0
Qwen Audio 3.1 TTS Next Alibaba Cloudseed0 to 2147483647
Qwen Audio 3.1 TTS Plus Alibaba Cloudseed0 to 65535; default 0
Qwen Image 2.0 Alibaba Cloudseed0 to 2147483647
Qwen Image 2.1 Alibaba Cloudseed0 to 2147483647
Qwen Image 3.0 Alibaba Cloudseed0 to 2147483647
Seedance 2.5 ByteDanceseed-1 to 2147483647
Seedance 2.5 Flash ByteDanceseed-1 to 2147483647
Stable Audio 2.0 Stability AIseed-
Stable Audio 2.5 Stability AIseed-
Stable Audio 3 Medium Stability AIseed0 to 2147483647
TRELLIS.2 4B Microsoftseed0 to 2147483647; default 42
TTS 1.5 Max Inworldtemperature0.1 to 2; default 1
TTS 1.5 Mini Inworldtemperature0.1 to 2; default 1
TTS 2 Inworldtemperature0.1 to 2; default 1
TTS 2 Flash Inworldtemperature0.1 to 2; default 1
Wan 2.1 Alibaba Cloudseed-
Wan 2.2 Alibaba Cloudseed-
Wan 2.5 Alibaba Cloudseed-
Wan 2.6 Alibaba Cloudseed-
Wan 2.7 Alibaba Cloudseed-
Wan 3.0 Alibaba Cloudseed0 to 2147483647
Wan2.1 Image Alibaba Cloudseed0 to 2147483647
Wan2.2 Image Alibaba Cloudseed0 to 2147483647
Wan2.5 Image Alibaba Cloudseed0 to 2147483647
Wan2.7 Image Alibaba Cloudseed0 to 2147483647
Whisper Large v3 Turbo OpenAItemperaturedefault 0,0.2,0.4,0.6,0.8,1

Frequently asked questions

What temperature should I use?

Use 0 to 0.3 for extraction, classification and code, and about 0.5 to 1.0 for conversation and writing. Check each model's accepted range, since some stop at 1.

Should I change temperature and top_p together?

Usually not. Adjust one of them and leave the other at its default, so you can see the effect of each change.

Does the same seed always give the same output?

Not always. On text models a seed makes output more repeatable, but the result can still differ slightly between runs, so store a response when you need it exactly.

Does temperature 0 make a model deterministic?

It makes output close to deterministic, because the model almost always picks its most likely token. Small differences between runs can still occur.

Ready to use better endpoints?

Check out our pricing or reach out if you want your own model deployed on our stack.