Skip to content
Frontier Leaf FL markFLFrontierLeaf

A field journal of new AI developments.

Model updates

Hugging Face releases Granite 4.1 LLMON preview models to separate instructions from data

Hugging Face released experimental Granite 4.1 models trained on LLMON markup, which benchmarks show can reduce indirect prompt injection success rates to nearly zero.

Depth
3bHugging Face released granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview as experimental models · reported
8bHugging Face released granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview as experimental models · reported
1 outlet · newest source 88h ago
Hugging Face releases Granite 4.1 LLMON preview models to separate instructions from data

Key points

  • Hugging Face released granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview as experimental models.
  • LLMON uses reserved special-token delimiters to explicitly mark structure, unlike JSON which relies on statistical patterns.
  • For some benchmarks, indirect prompt injection success rates drop to nearly zero for LLMON models compared to standard instruction-tuned models.

Hugging Face has released two experimental preview models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, which are trained to consume and generate text marked with LLMON. According to a blog post published by Hugging Face, this LLM-native markup language aims to address the brittleness of smaller models by explicitly communicating structure and semantics.Hugging Face has released two new experimental models that use a special markup language to help AI better understand the difference between instructions and data. The models are derived from the Granite 4.1 base and include a GitHub repository with toolkits for workflows.

Evidence: confirmedLLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B

aifleet-huggingface

Evidence: reportedLLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B

aifleet-huggingface

Evidence: reportedLLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B

aifleet-huggingface

The evidence

Hugging Face reported that indirect prompt injections hidden in user data are mostly ignored by LLMON models due to explicit instruction and data separation. For some benchmarks, attack success rates dropped to nearly zero, whereas regular instruction-tuned models were hijacked with high probability. These security improvements were achieved while maintaining general capability relative to Granite 4.1 Instruct models. The platform also noted that unrelated passages and old chat turns do not get pulled into answers because instructions are bound to specific data.

Evidence: reportedLLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B

aifleet-huggingface

Evidence: reportedLLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B

aifleet-huggingface

The decision

The introduction of LLMON suggests a shift away from relying on JSON for structural encoding in causal language models. Hugging Face explained that JSON was designed for parsers, not models, leading to ambiguity where quotation marks inside documents are treated the same as field closers. LLMON uses an XML-like markup with reserved special tokens, ensuring structure and content use separate parts of the vocabulary. This allows tags to be stated at the beginning of each span with explicit boundaries.

The takeaway

Developers can use conversion tools to translate existing JSON into LLMON without loss, according to the Hugging Face advisory. The material states that nesting depth is encoded locally in LLMON, so models do not need to track hierarchy over long spans, a known weakness for transformers. LLMON does not replace chat templates but enriches them. Teams can also define custom structures and tags, then tune a LoRA on top of the LLMON models to understand them.

Evidence: reportedLLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B

aifleet-huggingface

Building on this

Enterprise developers can build applications that leverage unique IDs to point to specific instructions anywhere in the context, binding them to data payloads analogous to function calls. This allows for clearer separation between system prompts, retrieved documents, and user data. The documented capabilities include the ability to ignore unrelated tool outputs and old chat turns, reducing noise in generated answers. Developers can integrate these models via Mellea, which has native LLMON support enabled by automatically LLMONizing its LLM calls.

Why it matters

  • Indirect prompt injection success rates drop to nearly zero in some benchmarks, offering a measurable security improvement over standard instruction-tuned models.
  • The explicit separation of instructions and data prevents unrelated passages and old chat turns from being incorrectly pulled into answers.
  • Conversion tools allow teams to migrate existing JSON pipelines to LLMON without loss, reducing the barrier to adoption for current workflows.
  • Custom structures can be defined and taught to the model via LoRA tuning, enabling flexible integration of proprietary data schemas.

Topics: Hugging Face, Granite, LLMON, Prompt Injection, Enterprise AI

Updated

Then / now / next

Then

LLMs relied on implicit context or JSON formats not designed for them, leading to brittleness and vulnerability to prompt injections.

Now

Hugging Face has released experimental Granite 4.1 models trained on LLMON, an LLM-native markup language that explicitly marks structure and semantics.

Next

Developers can test the granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview models and use the provided GitHub toolkit for workflows.

Story timeline

  1. Then LLMs relied on implicit context or JSON formats not designed for them, leading to brittleness and vulnerability to prompt injections.
  2. 12 Nov 2026 LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B
  3. Now Hugging Face has released experimental Granite 4.1 models trained on LLMON, an LLM-native markup language that explicitly marks structure and semantics.
  4. Next Developers can test the granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview models and use the provided GitHub toolkit for workflows.

The strongest doubt

Counterpoint from the sources

The models are described as experimental previews, implying they may not yet be stable for production use.

For readers

Developers

  • Use conversion tools to translate existing JSON to LLMON without loss.
  • Define custom structures and tune a LoRA on top of LLMON models to understand them.

Business

  • Indirect prompt injection success rates drop to nearly zero in some benchmarks.
  • General capability is maintained relative to Granite 4.1 Instruct models.

Researchers

  • LLMON addresses the weakness of transformers in tracking nesting depth over long spans.
  • Structure and content use separate parts of the vocabulary in LLMON.

Numbers in context

3breported8breportedCircle area is proportional to the magnitude; labels retain signs. Figures may describe different measures. Read their contexts above.

Ask the article

Local keyword search of this article; no generated answers.

Which of these is confirmed?

Check the answer

Hugging Face has released two experimental preview models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, which are trained to consume and generate text marked with LLMON.

The matched primary source and announcement wording support the confirmed label.

Did this article give you the evidence you needed?

No vote saved on this device.
Your choice stays on this device. No public tally.

Questions readers ask

What models are included in this release?

The release includes granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, which are experimental preview models derived from Granite 4.1 base models.

How does LLMON differ from JSON?

LLMON uses reserved special-token delimiters for structure, whereas JSON relies on characters like brackets and quotes that may appear in normal text, causing ambiguity for models.

Is LLMON a replacement for chat templates?

No, the source material states that LLMON does not replace chat templates but rather enriches them with explicit structural markup.

Connected stories

This storyllama.cpp v1.0.0 sets strict API stability for local AI developersllama.cpp v1.0.0 sets strict APIDatabricks Genie agent executes unsupervised multi-step tasks using platform ontologyDatabricks Genie agent executes Dartmouth provost resigns over undisclosed AI use in academic writingDartmouth provost resigns over uAWS VP Colleen Aubrey advises using existing business metrics to measure AI successAWS VP Colleen Aubrey advises us

Reading receipt

  • Hugging Face released granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview as experimental models.
  • LLMON uses reserved special-token delimiters to explicitly mark structure, unlike JSON which relies on statistical patterns.
  • For some benchmarks, indirect prompt injection success rates drop to nearly zero for LLMON models compared to standard instruction-tuned models.

Still open: What are the performance costs of converting existing JSON pipelines to LLMON in high-throughput environments?

Still open: How does the model handle malformed LLMON tags compared to malformed JSON?

Time spent: 0:00 · Depth: Full

Sources

  1. Hugging Face

Related stories

Background reading