LLMON: Markup for Training AI Models

Hugging Face released LLMON, an LLM-native markup language, in the latest Granite 4.1 3B and 8B models. The new markup lets developers annotate prompts and training data to guide AI agents toward specific outputs or behaviors.

LLMON is designed for use in prompting, few-shot learning, and fine-tuning of large language models. It provides a simple syntax for specifying prompts, labels, and instructions directly within the model's training data or inference pipeline.

What happened

On 2026-11-12, Hugging Face released LLMON, an LLM-native markup language, in the latest Granite 4.1 3B and 8B models. LLMON lets you define prompts and instructions for fine-tuning AI language models, built into the model itself rather than in a separate text file.

Granite is a suite of foundational models from IBM Research. The 3B and 8B refer to the number of parameters, or "size", of the models. Larger models are trained on more data and able to handle more complex tasks.

Hugging Face says LLMON is designed to make it easier to share and reuse prompts and instructions between different model architectures and fine-tuning runs. It can be used with any Hugging Face model, not just Granite.

markdown

Why it matters

Hugging Face's LLMON opens new ways to fine-tune and customize large language models for specific domains and applications. The markup language lets developers express domain knowledge and constraints in a structured way directly in the model. This could make customizing AI more accessible and less error-prone than current methods, which often require hand-editing weights.

LLMON also enables new classes of pre-trained models that can be quickly adapted for different purposes with minimal fine-tuning. This could lower the bar to building domain-specific applications that leverage the power of large language models. The approach could accelerate AI adoption in enterprise software, customer service, and other domains where fine-tuned text understanding is needed.

What to do

• Adopt LLMON for your in-house LLM training projects, particularly those using IBM's Granite models.
• Run the LLMON samples and documentation to familiarize yourself with the new syntax and capabilities.
• Update existing LLM training pipelines to use LLMON markup where appropriate to improve maintainability and code quality.
• Consider integrating LLMON into your AI/ML training curriculum or courses, as it will be an increasingly important skill for data scientists and engineers.
• Analyze whether your current LLM training practices could benefit from the modular approach provided by LLMON, and make plans to migrate any relevant projects.

2TI lens

Markup for training AI models is one way we help make AI more accessible and productive at our clients. When the work is clear, code is faster and cheaper to write and that means better results for less cost. We map out each project's AI scope and tasks, then use markup like LLMON to accelerate deployment of production AI. This lets us iterate more quickly and deliver more value with less engineering.

Use a list of source names from the source pack and a few sentences for each. Use the provided output template. Use a short paragraph with no markdown fence. Use a

• list for the sources. Use

  for a short paragraph. Use

  Sources

  for the section title.

  2TI

  Hugging Face

  Google

  Hugging Face

  Hugging Face

  Hugging Face

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google

  Google