IBM releases two Granite previews that read LLMON markup

IBM has released two experimental preview models trained to read and write LLMON, a markup language for the text a model sees. The previews are granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, at 3 billion and 8 billion parameters. Both start from Granite 4.1 base models, not the released instruct models.

The markup marks instructions, data, and requirements as separate blocks inside a model's context. LLMON stands for LLM Object Notation, and the page says it is pronounced lemon. Tags sit in reserved special tokens, so the structure is not written as ordinary characters. A unique ID can point at an instruction and bind it to the data it should use.

The aim is to stop a hidden instruction in user data from being treated as a command. On some tests, those hidden instructions are mostly ignored, and attack success falls to nearly zero, while ordinary instruct models are hijacked at a high rate. Across standard benchmark groups, the page says the previews stay in line with Granite 4.1 instruct models.

LLMON does not replace a chat template. A chat template still marks who is speaking. LLMON adds what each piece means, and how the pieces relate, inside and across those turns. Teams that already store data as JSON can convert it to LLMON and back without losing the structure.

Training starts from the base checkpoints and rebuilds an instruct set in the new markup. Extra blocks that look like instructions are left unbound, so the model is taught to follow only the block that names the task. Later training includes five reinforcement stages and a final preference stage on a thousand pairs drawn from the model's own replies to injection attempts.

The page calls that separation a learned behavior, not a guarantee. A toolkit and examples are on GitHub. The page also shows a small adapter, tuned on ScienceWorld, that uses custom tags for a science-task agent.