IBM has released two experimental preview models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, trained to read and generate LLMON, a markup language that makes the structure of an LLM's context explicit. Pronounced “lemon,” LLMON wraps components such as system prompts, instructions, data and requirements in reserved special-token delimiters, and uses unique IDs to bind each instruction to the data it should act on. IBM argues that JSON, which many developers currently use to structure prompts, was designed for parsers rather than language models. In its view, quotation marks and brackets that close fields look the same to a model as those appearing in ordinary text, and nesting depth is hard for transformers to track over long spans.
The main claimed benefit is resistance to indirect prompt injection, in which malicious instructions hidden in user data hijack a model. IBM says attack success rates on some benchmarks fall to nearly zero, compared with high rates for regular instruction-tuned models, while general capability is maintained relative to the Granite 4.1 Instruct models. LLMON does not replace chat templates. The preview models ship with their own template, so developers pass the usual messages list and mark individual parts as executable or not. IBM has also published a GitHub toolkit with converters that translate JSON to LLMON and back, along with example workflows and templates. The Mellea framework supports LLMON natively and applies it automatically to its model calls.
The release is explicitly experimental, and the benchmark results come from IBM's own testing. How the approach performs in production systems, and against a wider range of attacks, has not yet been shown by independent parties. For Australian businesses and professionals, the change is not immediate. The models are previews, and the material does not describe pricing, licensing for commercial use or production support. Teams building tools that process untrusted documents, emails or customer submissions may find the prompt-injection results worth testing in a sandbox, but should not treat them as a security control. Any testing involving personal information still falls under the Privacy Act 1988 and the Australian Privacy Principles, so de-identified data is the safer choice.
Key Points
IBM released two experimental Granite 4.1-based preview models, 3B and 8B, trained to consume and generate LLMON.
LLMON uses reserved special-token delimiters and unique IDs to separate instructions from data and bind them together.
IBM reports attack success rates on some prompt-injection benchmarks falling to nearly zero, with general capability maintained.
A GitHub toolkit provides lossless converters between JSON and LLMON, and Mellea supports LLMON natively.
The models are experimental previews, and the results are IBM's own, so independent validation is still needed.