IBM
Granite 4
downloadable enterprise models and structured-prompt previews

Key facts
- IBMGranite family
- Maker
- Granite 4.2reasoning model, 25 Aug 2026
- Generation
- Open weightdownloadable
- Licence
- 12 Nov 20263B and 8B model classes
- LLMON previews
- Business applicationsdeployment and testing
- Built for
IBM Granite is a family of downloadable AI models. The Granite 4.1 LLMON previews use structured prompt blocks to separate instructions from data.
Granite models can run on your own infrastructure
IBM Granite is a family of downloadable AI models for business applications. The official Granite repositories provide checkpoints, licences and deployment instructions. Each model has its own inputs, limits and intended uses, so teams should choose a particular checkpoint before evaluating it.
Granite 4.2 adds native reasoning
IBM released Granite 4.2-8B on 25 August 2026. This Apache 2.0 model adds native reasoning to the family. It is a separate checkpoint from the Granite 4.1 LLMON previews released in November.
LLMON previews add structured prompt blocks
IBM introduced Granite 4.1 LLMON previews on 12 November 2026. The 3B preview and 8B preview accept and return text, list 12 languages and use Apache 2.0.
LLMON is a markup format that gives instructions and supplied data explicit structure. The previews are trained to use that structure. It can help the model distinguish a developer’s instructions from material being analysed, but applications still need permission controls and testing against malicious input.
The November release concerns these preview variants. It is separate from the original Granite 4.1 family release dated 29 April 2026 in this catalogue.
Evaluate the checkpoint and the surrounding application
Open weights let a team operate and adapt the model within its own infrastructure, subject to the licence terms. Hardware requirements depend on model size, numerical format, context length and runtime. Hosting and governance tools are separate parts of that deployment.
For a useful assessment, test representative documents, tool calls and failure cases. Check how the model handles conflicting instructions and data from untrusted sources. A model licence or a provider’s governance certification does not by itself establish that a particular application meets a regulatory requirement.
See the model security capability board for evaluations with their sources and test conditions, and the LLM directory for other model families.
More in Large Language Models
All LLMs →- Google DeepMindGemini 4 Argonlonger reasoning, coding and cyber defence
- OpenAIGPT-5.6the flagship from 9 July to 3 September 2026
- OpenAIGPT-6 Astrathe new generation, rolling out from 3 September 2026
- AnthropicClaude Fable 5.2An AI model for difficult coding, research and long-running tasks
- Google DeepMindGemini 4 CarbonCoding, research and tasks with many steps, now available more widely
- OpenAIGPT-6.1 Solthe upgraded mid-price GPT-6 model, released 29 September 2026