We use temperature 0 for 15 of the 29 LLMs and document the settings for the other 14.
We use temperature 0 with Gemini 3.1 Pro Preview and Gemini 3.5 Flash, below the default of 1.0 in the Gemini 3 developer guide.
We use medium reasoning effort with GPT-5.5, the default of our connector, and the three GPT-5.6 models at low effort. From GPT-5.1, OpenAI reasoning models accept temperature only at reasoning effort none, so at low effort we translate with the GPT-5.6 models without temperature control.
Sampling settings by model
| Model | Vendor | Type | Temperature | Reasoning setting | Note |
|---|---|---|---|---|---|
| Temperature 0 (15 models) | |||||
| Jamba Large 1.7 | AI21 | LLM | 0 | ||
| Claude Haiku 4.5 | Anthropic | LLM | 0 | ||
| Claude Sonnet 4.6 | Anthropic | LLM | 0 | ||
| Command A Translate | Cohere | LLM | 0 | ||
| Tiny Aya | Cohere | LLM | 0 | ||
| Gemini 3.1 Pro Preview | LLM | 0 | |||
| Gemini 3.5 Flash | LLM | 0 | |||
| M2.5 | MiniMax | LLM | 0 | ||
| Large 3 | Mistral | LLM | 0 | ||
| Medium 3.1 | Mistral | LLM | 0 | ||
| GPT-4.1 | OpenAI | LLM | 0 | ||
| Grok 4.20 (non-reasoning) | xAI | LLM | 0 | ||
| Grok 4.20 (reasoning) | xAI | LLM | 0 | ||
| GLM-5 | z.ai | LLM | 0 | ||
| Qwen3-235B-A22B | Alibaba | open-weight LLM | 0 | ||
| A fixed value other than 0 (6 models) | |||||
| ALL-Lab | ALL-Lab | MT engine | creativity 0.7 | A creativity setting instead of temperature, range 0.1 to 1.0, left at its default. | |
| Kimi K2.5 | MoonshotAI | LLM | 1 | The model requires temperature 1. | |
| Llama 4 Maverick | Meta | open-weight LLM | 0.1 | The default of our connector. | |
| gpt-oss-120b | OpenAI | open-weight LLM | 0.1 | effort low | The default of our connector. |
| Bielik-11B-v3.0-Instruct | SpeakLeash | open-weight LLM | 0.1 | The default of our connector. | |
| Apertus-8B-2509 | Swiss AI | open-weight LLM | 0.1 | The default of our connector. | |
| Vendor defaults: no temperature in the request (9 models) | |||||
| Claude Fable 5 | Anthropic | LLM | effort low | The API rejects temperature with an error. | |
| Claude Opus 4.8 | Anthropic | LLM | thinking off | The API rejects temperature with an error. | |
| Claude Sonnet 5 | Anthropic | LLM | effort low | The API rejects temperature with an error. | |
| V3.2 | DeepSeek | LLM | Reasoning model: the API does not accept temperature. | ||
| GPT-5.5 | OpenAI | LLM | effort medium | Reasoning model: the API does not accept temperature. | |
| GPT-5.6 Luna | OpenAI | LLM | effort low | Reasoning model: the API does not accept temperature. | |
| GPT-5.6 Sol | OpenAI | LLM | effort low | Reasoning model: the API does not accept temperature. | |
| GPT-5.6 Terra | OpenAI | LLM | effort low | Reasoning model: the API does not accept temperature. | |
| Gemma 4 | open-weight LLM | The API has no temperature setting. | |||
| No sampling parameter: MT engines (27 models) | |||||
| Machine Translation API | Alibaba | MT engine | |||
| Translate | Amazon | MT engine | |||
| NMT | AppTek | MT engine | |||
| Translate API | Baidu | MT engine | |||
| Machine Translation System | Elia | MT engine | |||
| Cloud Advanced Translation API | MT engine | ||||
| MT | iFlyTek | MT engine | |||
| Translator | Microsoft | MT engine | |||
| Papago NMT Commercial | Naver | MT engine | |||
| NiuTrans | NiuTrans | MT engine | |||
| Cloud Language Service | Oracle | MT engine | |||
| PangeaMT | Pangeanic | MT engine | |||
| PNMT API | SYSTRAN | MT engine | |||
| NMT | Tarjama | MT engine | |||
| TartuNLP | TartuNLP | MT engine | |||
| Cloud TMT | Tencent | MT engine | |||
| ModernMT Enterprise Edition | Translated | MT engine | |||
| Cloud Translate v2 | Yandex | MT engine | |||
| Translation Service API | Youdao | MT engine | |||
| Next Gen | DeepL | LLM-based MT engine | |||
| Translation LLM | LLM-based MT engine | ||||
| LLM | Supertext | LLM-based MT engine | |||
| TildeOpen LLM engine | Tilde | LLM-based MT engine | |||
| Lara 2 | Translated | LLM-based MT engine | |||
| Widn AI Tower Anthill | TransPerfect | LLM-based MT engine | |||
| Widn AI Tower Sugarloaf | TransPerfect | LLM-based MT engine | |||
| Widn AI Tower Vesuvius | TransPerfect | LLM-based MT engine | |||
Temperature: the value in our requests. Reasoning setting: reasoning effort or extended thinking, for the models that have them.