LLM
Introduction
The connection to an LLM is based on
- An API protocol (OpenAI or specific)
- A URL endpoint, as shown below
- An API token the customer must enter when creating the configuration.
For example, to obtain your key from ChatGPT, please refer to the following page:
Starting with 5.0, an LLM connection can be used for translation, for AI Quality Review (scoring source and target strings), or both. See LLM Tasks below.
Example of LLM Types and Models
| LLM | API Endpoint URL | Model Examples |
|---|---|---|
| ChatGPT (OpenAI) | https://api.openai.com/v1/chat/completions |
gpt-5.5, gpt-5.4-mini, o3 |
| Claude (Claude) | https://api.anthropic.com/v1/messages |
claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5 |
| DeepSeek (OpenAI) | https://api.deepseek.com/v1/chat/completions |
deepseek-v4-pro, deepseek-v4-flash, deepseek-r1 |
| Fireworks (OpenAI) | https://api.fireworks.ai/inference/v1/chat/completions |
accounts/fireworks/models/llama-v3p3-70b-instruct, accounts/fireworks/models/deepseek-v3 |
| Gemini (Gemini) | https://generativelanguage.googleapis.com/v1beta/models |
gemini-3.1-pro, gemini-3.5-flash, gemini-2.5-flash-lite |
| Groq (OpenAI) | https://api.groq.com/openai/v1/chat/completions |
llama-3.3-70b-versatile, gpt-oss-120b |
| Mistral (OpenAI) | https://api.mistral.ai/v1/chat/completions |
mistral-large-3, mistral-small-4, codestral |
| OpenRouter (OpenAI) | https://openrouter.ai/api/v1/chat/completions |
openai/gpt-5.5, anthropic/claude-sonnet-4-6, google/gemini-3.1-pro |
| Perplexity (OpenAI) | https://api.perplexity.ai/chat/completions |
sonar-pro, sonar |
| Together AI (OpenAI) | https://api.together.xyz/v1/chat/completions |
meta-llama/Llama-3.3-70B-Instruct-Turbo, deepseek-ai/DeepSeek-V4-Pro |
| xAI Grok (OpenAI) | https://api.x.ai/v1/chat/completions |
grok-4, grok-4-fast, grok-3-mini |
Note: European customers may lean towards Mistral.
The LLM models change often. The above are examples as of June 2026.
LLM Creation Information
Below is a view of the Create page.
Pick OpenAI for all LLM whose API is compatible, that's all the LLM but Claude and Gemini. In the example below, Mistral uses the OpenAI protocol; Please note the Mistral URL.
LLM Tasks: Translation and Quality Review (Upcoming in 5.0)
Localyzer uses LLMs for two separate tasks:
| Task | What it does | Prompt |
|---|---|---|
| Translation | Translates source strings into the target locales. | Translation prompts: connection, project, and string level, plus context images. |
| Quality Review | Scores source strings for correctness and target strings for translation quality, from 1 to 5 (5 is best). See AI Quality Review. | Quality review prompt |
Each task is configured independently: it can use its own LLM connection, model, and prompt, and the selection can be different for each project.
Recommendations:
- Use a different model for quality review than the one used for translation, so the translation is evaluated independently.
- Match the model to the task to save tokens and cost, for instance a smaller model for one task and a more capable model for the other.
- If strings have context images, choose models that accept images for both tasks.
File:LLM Translation Quality Selection.png
Token Count
When viewing the LLM connection, the number of tokens used up to that point will be displayed.
When an LLM is used for translation or quality review, the number of tokens will be shown in the log file for that task. The total tokens used for the project will be tracked on the Localyzer page.
Prompting
For best results with an LLM, prompting is recommended. Localyzer allows for three levels of prompting for translation, and a separate prompt for quality review:
See: Prompting
Context Prompting
If the LLM connection is configured with an InContext server, then when a context has been associated with a string (See String Manager / Context), the context will be passed to the LLM to provide the necessary information to better translate that string, and to better evaluate its translation during quality review.
Make sure the LLM model takes advantage of the InContext image. Not all models do.