Hugging Face releases Granite 4.1 previews with LLMON language
🎧 Listen to this story
Hugging Face has released the experimental preview models granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, derived from Granite 4.1 base models. The company also published a GitHub repository with tools, examples, JSON-to-LLMON converters and support for use with Mellea.
LLMON uses special-token delimiters and tags to represent instructions, data, requirements and nested structures explicitly for language models. Hugging Face says this brought indirect prompt-injection success rates down to nearly zero on some benchmarks, while instruction-tuned models were hijacked with high probability. The company says general capability was maintained relative to Granite 4.1 Instruct models and explains that LLMON complements, rather than replaces, chat templates.
Why it matters · editorial interpretation
If the results cited by Hugging Face hold beyond the mentioned benchmarks, explicitly representing instructions and data could offer an additional approach to reducing prompt-injection hijacking. The published repository also lets developers test the proposal and its integration with Mellea.
Sources
Security