Skip to content
Main Site News Console

LLMON: LLM Object Notation, Now Integrated into Granite 4.1 3B and 8B

· Hugging Face Translated
教程模型卡

Published on November 12, 2026

Large language models (LLMs) rely on their context to generate the next set of tokens. Over long interaction trajectories, the context fills up with diverse types of components: the system prompt, previous conversation turns, instructions operating on retrieved documents, tool responses or user data, requirements restricting the range of permitted outputs, and chains of reasoning. The context carries rich structure, yet much of it is either implicit to the model or encoded in formats like JSON that were not designed for LLM consumption. This is one of the primary reasons why LLMs—especially smaller-scale models—can be brittle and sometimes fail unexpectedly. The lack of explicit structure can also be exploited by attackers, for instance, through prompt injections.

To address this, we developed LLMON (LLM Object Notation, pronounced “lemon”), an LLM-native markup language designed to explicitly convey structural and semantic information to the model. Semantically coherent components are labeled with meaningful tags and enclosed within reserved special-token delimiters. Data schemas (including nested structures) are declared in a way that is completely unambiguous to a causal LLM. Using unique IDs, the model can point to specific instructions located anywhere in the context and bind them to the data payloads they are meant to process—similar to how a function is called and passed arguments in a programming language.

Today, we are releasing two experimental preview models derived from the Granite 4.1 base models, specifically trained to understand and generate LLMON-marked text: granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview. Alongside the models, we are open-sourcing a GitHub repository containing a toolkit for working with LLMON, as well as sample workflows and templates.

*What it enables: