The Pulse
IBM releases Granite 4.1 3B LLMON model in preview
IBM has released a 3B-parameter Granite preview trained to use a structured prompt format called LLMON. The company says the experimental model is for evaluation, not production, and that its instruction boundaries are not guaranteed.

IBM's Granite 4.1 3B LLMON preview is a 3-billion-parameter instruction model trained to read a structured prompt format instead of a conventional chat template. The model card lists its release date as November 12, 2026, and its license as Apache 2.0.
Source huggingface.co
IBM also released an experimental 8B LLMON preview. The company describes both as models for exploration and evaluation, not production.
How the prompt format works
LLMON, short for LLM Object Notation, assigns individually addressable blocks to components such as instructions, documents, data, tool definitions and tool results. An execution block specifies which instruction to run, which inputs it uses and where the output belongs.
The model includes a chat template that renders messages into the format, so developers do not normally need to write the markup by hand. The format uses six reserved delimiters, each represented by a single token, while block identifiers distinguish components within a document.
IBM says its training data included decoy instructions and data blocks that appeared in the context but were deliberately left unbound to the execution block. Its training process also included five sequential reinforcement-learning stages and a final direct preference optimization stage. The latter used a thousand on-policy preference pairs drawn from the model's own responses to injection attempts.
Injection results and limits
In an IBM evaluation on the OPI benchmark, the LLMON previews avoided 47% more attacks relative to the best frontier model tested and 95% more than the best open-source model tested. The benchmark places indirect prompt injections in data used for natural-language-processing tasks; the reported measure was the share of attacks in which the model did not carry out the injected instruction.
Those benchmarks test static indirect prompt injections. The company says it is also working to improve performance against adaptive attacks that unfold over many chat turns.
IBM cautions that the rule separating instructions from input is learned behavior, not an enforced guarantee: an instruction embedded in a document, data or tool result can still be followed. It recommends against treating the model as the sole defense for an agent with access to sensitive systems.
Benchmarks and intended uses
The model card reports a 68.05 score for the preview on the five-shot MMLU benchmark, compared with 67.02 for the conventionally formatted Granite 4.1 3B instruct model. On MMLU-Pro, using five-shot chain-of-thought evaluation, the preview scored 46.08 against 49.83 for the instruct model. Those general-capability tests do not measure whether the model maintains the instruction-and-data boundary when content is unbound or adversarial.
The model is intended for general instruction following, including as a foundation for assistants and agents with tool-use capabilities. Listed capabilities include summarization, text classification, extraction, question answering, retrieval-augmented generation, code-related tasks and function calling.
The card lists English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch and Chinese as supported languages. It also warns that conventional chat-format prompts are out of distribution for the model and will degrade quality; the bundled template renders input in the expected format.
Like its base model, the preview may produce inaccurate, biased or otherwise objectionable content, according to the card. The structured prompt format does not remove that risk.