IBM released two experimental Granite 4.1 models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, designed to consume and generate LLMON. LLMON is a markup language built to explicitly communicate structure to large language models.
The models use reserved special-token delimiters to separate structural tags from content, unlike JSON which uses characters. IBM claims the new Granite 4.1 LLMON models show 47% more resistance to indirect prompt injection attacks compared to the best frontier model tested.
Performance on standard benchmarks remains on par with the base Granite 4.1 Instruct models. A GitHub repository and a toolkit for working with LLMON, including JSON conversion tools, are available for developers to test the new architecture.
