Key points
- IBM released granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview models.
- The LLMON markup language separates instructions from data using reserved special tokens.
- Benchmarks show near-zero attack success rates for indirect prompt injections.
IBM published a blog post announcing two experimental preview models derived from Granite 4.1 base versions, trained to consume and generate LLMON-marked text.confirmed The release includes a GitHub repository with a toolkit for working with the new markup language.IBM released new models that use a special markup language to make AI outputs more secure and structured.Granite 4.1 preview models now process LLMON, an XML-like markup using reserved special tokens to separate structure from content.
Evidence: confirmed
LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8Baifleet-huggingface
The claim
The company states that LLMON addresses brittleness in smaller models by explicitly communicating structure and semantics. By wrapping semantically coherent components in meaningful tags with reserved special-token delimiters, the system makes data schemas unambiguous to causal language models. Unique IDs allow specific instructions to bind to data payloads, similar to function calls in programming.
The response
According to the blog post, indirect prompt injections hidden in user data are mostly ignored by these models due to explicit instruction and data separation.reported Benchmarks indicate that attack success rates drop to nearly zero, whereas regular instruction-tuned models are hijacked with high probability. This security improvement occurs while maintaining general capability relative to standard Granite 4.1 Instruct models.reported
Evidence: reported
LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8Baifleet-huggingface
What the evidence means
The material suggests that traditional JSON formats are ill-suited for causal language models because tokenizers do not respect grammar rules. Structure in JSON is encoded by tokens that also appear in normal text, making escaping rules statistical patterns rather than clear boundaries. LLMON uses an XML-like markup with reserved special tokens, separating structure from content in the vocabulary.
In practice
Enterprise IT buyers can use provided converters to turn existing JSON into LLMON without loss. Teams can define custom structures and associated tags, then tune a LoRA on top of the LLMON models to understand them. The system enriches rather than replaces chat templates. Mellea offers native LLMON support by automatically LLMONizing LLM calls for users who want to try the technology immediately.
Why it matters
- Reduced prompt injection risk addresses a critical security concern for enterprise deployments of smaller models.
- Explicit structure binding prevents unrelated passages or old chat turns from influencing answers.
- Lossless JSON conversion tools allow teams to integrate LLMON into existing pipelines without rewriting data formats.
- Custom tag definitions via LoRA tuning enable organizations to adapt the markup to specific application requirements.
Topics: models, IBM, Hugging Face, security, markup language
Updated
Then / now / next
LLMs relied on implicit context structures or JSON formats not designed for model digestion.
IBM released preview models trained to process explicit LLMON markup with reserved special tokens.
Teams may adopt LLMON to mitigate prompt injection risks in smaller-scale deployments.
Story timeline
- Then LLMs relied on implicit context structures or JSON formats not designed for model digestion.
- 12 Nov 2026 LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B
- Now IBM released preview models trained to process explicit LLMON markup with reserved special tokens.
- Next Teams may adopt LLMON to mitigate prompt injection risks in smaller-scale deployments.
The strongest doubt
Counterpoint from the sources
The models are experimental previews, and general capability is maintained relative to standard Instruct models, but specific performance trade-offs for complex tasks are not detailed.
Numbers in context
Circle area is proportional to the magnitude; labels retain signs. Figures may describe different measures. Read their contexts above.Try it
Convert this JSON to LLMON
Good result: The converter tool translates the input into LLMON format without data loss.
Ask the article
Local keyword search of this article; no generated answers.Which of these is confirmed?
Check the answer
IBM published a blog post announcing two experimental preview models derived from Granite 4.1 base versions, trained to consume and generate LLMON-marked text.
The matched primary source and announcement wording support the confirmed label.
Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
What models support LLMON?
The granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview models support the markup language.
Does LLMON replace chat templates?
No, LLMON enriches chat templates rather than replacing them entirely.
How does LLMON handle JSON data?
Conversion tools translate JSON to LLMON and back in a lossless manner.
Connected stories
Reading receipt
- IBM released granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview models.
- The LLMON markup language separates instructions from data using reserved special tokens.
- Benchmarks show near-zero attack success rates for indirect prompt injections.
Still open: How does LLMON performance compare to larger non-LLMON models on complex reasoning tasks?
Still open: What are the computational costs of converting large datasets to LLMON format?