LLM Chat

Overview

LLM Chat puts open-weight chat models hosted on Wiro behind a single input form. The base model is one more field in the request, so switching from a 1.5B model to a 32B one, or sending the same prompt to two models, leaves everything else as it was. The list spans Llama 2 through Llama 3.2, Qwen from 1.5 to 3.8, Mistral 7B and Mistral Nemo, Gemma 2, Phi-3.5 and Phi-4, GPT-OSS 20B, the DeepSeek R1 distills, EuroLLM and Wiro AI's two Turkish models. Private chat models on your account show up in a separate list.

What you can build

• A chatbot that can move to a different base model later with the same request format
• Side-by-side tests that send one prompt with a fixed seed to several base models
• Assistants for many end users, each with its own conversation history
• Routing short requests to a small model and harder ones to a larger one

Inputs

• Base model: one public chat model, or one of your private models.
• Prompt (required), plus an optional system prompt that is prepended to steer tone and role.
• user_id and session_id, both optional. History is stored under the user_id; keep the same session_id to continue a thread, or send a new one to start a separate thread for that user.
• Sampling: temperature from 0 to 5 (default 0.7), top_p 0.95, top_k, repetition and length penalties, a seed (default 123456) and a do_sample switch.
• Length and stopping: at the default max_tokens of 0, the reply budget is 2,048 tokens, or twice the prompt length once the prompt runs past 2,048 tokens. A value you set is used as given, up to the room left in the model's window. min_tokens sets a floor, and stop sequences go in a semicolon-separated list.
• Quantization, on by default.

Outputs

The selected model's reply, as text.

Recommended settings

For translation or code, switch do_sample off so the same input returns the same reply. Keep the seed fixed while comparing base models. The page's quantization note says the quantized model may lose a little output quality, so if a reply looks wrong, run the same prompt with quantization disabled.

Limitations

Input length and language coverage depend on the base model you choose, so check that model's own page before sending long documents or non-English prompts.