IBM Granite is a family of open AI models built for enterprise use. The page describes language, vision, speech, guardrail, embedding and time series models, all released under the Apache 2.0 license.
The models are aimed at businesses that want to build and scale AI work with customizable models and flexible deployment. The page says they can run across any infrastructure, including on-premises, which it highlights as a use for IBM Granite 4.2.
The language models are described as dense, thinking models that the page says are competitive with larger models on a range of enterprise tasks. The page says they excel at reasoning and agentic tasks, from math and code problems to multi-step tool use, so teams can build AI-driven workflows.
The speech models are described for transcription across accents, domains and noisy environments. The vision models are described for understanding documents, charts and images. The Guardian models are guardrails for detecting harmful content and outputs. The embedding models support retrieval, search and classification. The time series models are described for forecasting and for running on a laptop.
The page says the models come with transparency into training data and processes, and it lists built-in guardrails. It also links to benchmark tables, a set of developer recipes, and a list of places to access the models, including Hugging Face, Ollama, LM Studio, watsonx.ai, OpenRouter and Replicate.
The page does not state pricing for Granite itself. The models are available to download, and the page links to IBM watsonx products for building and deploying applications on top of them.
