Chat Templates

Reviewed by Proof · Updated

Definition

A chat template is the fixed text format a model uses to turn a list of chat messages into the single token sequence it was trained on.

Key takeaways

  • Under the hood, a model reads one long string, not a list of messages.
  • The chat template adds special markers that show where each system, user and assistant message starts and ends.
  • Each model family has its own template. Using the wrong one can make a model behave badly.
  • When you call an API like luv13, the server applies the template for you. You just send messages.
  • You deal with templates directly only when you run Open-Weight Models yourself or use raw text completion.

What a template does

Say you send:

[
  {"role": "system", "content": "Be brief."},
  {"role": "user", "content": "Hi!"}
]

The server turns that into one string with role markers and special tokens, then adds the marker that means "the assistant speaks next". The exact markers differ by model family. That's the whole job of the template.

Templates also decide how tool definitions and tool results are laid out, which is one reason Tool Calling quality varies between models.

When it matters to you

  • Running models locally. Libraries such as Hugging Face Transformers store the template with the model and apply it with apply_chat_template. Local servers usually do it for you too.
  • Odd output. If a self-hosted model repeats role names or never stops, the template is a common cause.
  • APIs. With luv13 you never write a template. Send OpenAI-style messages to Chat Completions and luv13 handles the format.

Example

curl https://api.luv13.ai/v1/chat/completions \
  -H "Authorization: Bearer $LUV13_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "luv13/glm-5.3-flash",
    "messages": [
      {"role": "system", "content": "Be brief."},
      {"role": "user", "content": "Hi!"}
    ]
  }'