What is Granite?
Granite is IBM's family of open-weight language models, released under the permissive Apache 2.0 license. They are built for enterprise use: efficient sizes that run cheaply, code models, and a focus on transparency, governance, and data provenance rather than chasing the top of general leaderboards.
You use Granite through IBM's watsonx platform, by self-hosting the open weights (available on common hubs and runtimes like Hugging Face and Ollama), or through cloud marketplaces. IBM documents the training data and publishes the models for inspection.
Granite is the family to evaluate when you want open, permissively licensed models with an enterprise and compliance story, especially in regulated industries where provenance and governance matter as much as raw capability.
What it's best for
- Regulated and enterprise environments that need open, permissively licensed models with documented provenance.
- Efficient, cost-effective sizes for high-volume tasks and on-premises deployment.
- Code generation and software tasks through the Granite code models.
- Self-hosting under Apache 2.0, with no per-token vendor bill and full data control.
- Retrieval-augmented generation and tool use inside business applications, including through watsonx.
Where it falls short
- Being the single strongest model on open-ended general benchmarks; Granite trades peak capability for efficiency, openness, and governance.
- A consumer chatbot experience; the family is aimed at builders and enterprises.
- Native image, audio, or video generation; Granite is focused on text and code (with some specialized models).
Ways in
Self-host the open weights from common hubs and runtimes (for example Hugging Face or Ollama) for full control, or use the models on IBM's watsonx platform with its enterprise tooling.
Pick a Granite size that matches your hardware and latency budget; the smaller models are designed to run cheaply.
Enterprise and governance
IBM publishes details about the training data and licenses Granite under Apache 2.0, which matters for legal review and for industries that require provenance and the right to deploy freely.
On watsonx, combine Granite with retrieval, tuning, and governance features to ground answers in your own data and keep an audit trail.
Getting better answers
For grounded answers, pass the retrieved documents and ask the model to answer only from them and cite sources.
Use the Granite code models for software tasks, and fine-tune a smaller model on your own data when a specialized, efficient model beats a larger general one.
What Granite costs
Approximate, in USD, as of January 2026. Prices change often. Confirm on the official site before you rely on them.
Open weights
$0 (self-host)
Apache 2.0 licensed; download and run freely, paying only your own compute.
watsonx
Usage-based
Hosted on IBM's platform with enterprise tooling; priced by usage.
Enterprise
Custom
IBM enterprise agreements, support, and deployment services.
Example prompts
Copy these into Granite as starting points, then adapt them to your task.
Answer using only the retrieved internal documents. Cite the source for each claim, and if the documents do not answer the question, say so rather than guessing.
Using a Granite code model, write a function that does the following, with input validation and a short docstring. Explain any assumptions you made.
Extract the required fields from each document as JSON. Return only valid JSON and do not infer values that are not present; mark missing fields as null.
Summarize why an Apache 2.0 licensed, open-weight model like Granite can be easier to approve for a regulated deployment than a closed hosted model, and what we still need to verify.
Granite
common questions.
Direct answers to the questions we get asked the most. If yours isn't covered, write to the team.