IBM and Granite: open models built for the enterprise

A computing pioneer focused on business

IBM, headquartered in Armonk, New York, is one of the oldest names in computing. In AI history it is known for Deep Blue, which beat Garry Kasparov at chess in 1997, and later for Watson. Today its AI strategy targets businesses first, with the watsonx platform and an in-house model family: Granite.

A family built for specific jobs

Rather than a single giant model, Granite is a set of compact models, each with its own specialty:

• Granite (language): general-purpose large language models, whose latest versions can reason before answering, with a thinking mode you can switch on or off;
• Granite Vision: understanding documents, charts and images;
• Granite Speech: speech transcription;
• Granite Embedding: embeddings for search and RAG;
• Granite Guardian: guardrails that detect harmful content;
• Granite Time Series: time-series forecasting.

This split reflects a core IBM belief: in business, a smaller model that is well suited to its task costs less and is easier to control.

Openness and trust as selling points

Every Granite model is released as open weights under the Apache 2.0 license. IBM adds guarantees that are rare in the industry: the management system for its language models is certified to ISO 42001 (the AI management standard), and the models are cryptographically signed, so you can check that a downloaded file has not been tampered with. For a legal or compliance team, these points often matter more than a benchmark score.

How to access Granite

You can download the models from Hugging Face (the ibm-granite organization), run them locally with Ollama or LM Studio, or use hosted services such as watsonx.ai, OpenRouter or Replicate. The official documentation describes each version and how to use it.

To see where Granite stands among other open models, visit our open-source models page. If you are looking for another enterprise-focused lab, the Cohere profile describes a comparable approach.