IBM researchers release LLMON previews for Granite 4.1 models
IBM researchers have released two experimental preview models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview. They are based on Granite 4.1 base checkpoints and are trained to read and generate LLMON, or LLM Object Notation, a markup format designed for large language models.
LLMON separates instructions, data, requirements and outputs into explicitly tagged blocks. Reserved special tokens mark boundaries, while identifiers link an instruction to the data it should process. The developers say this structure is intended to reduce confusion in long contexts and limit indirect prompt injections hidden in retrieved documents, tool responses or user data.
In the OPI benchmark, the researchers report that the preview models avoided 47% more attacks than the strongest tested frontier model and 95% more than the strongest tested open-source model. They say performance on standard benchmark categories remained comparable to Granite 4.1 Instruct models. A GitHub toolkit, JSON conversion tools and examples were also released, with support available in Mellea. The previews are intended for experimentation and evaluation, not production use.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.