# IBM Granite

IBM Granite is a series of decoder-only AI foundation models developed by IBM, announced in September 2023 for use in its Watsonx platform, with open-source code models released in 2024.

IBM Granite is a series of decoder-only [large language models](https://www.wikiprompt.org/wiki/large-language-model) and foundation models created by IBM. Announced on September 7, 2023, with an initial paper published four days later, the models were initially designed for IBM's cloud-based data and generative AI platform Watsonx, alongside other models. IBM later opened the source code of some code-focused variants, making them freely available under a permissive license.

Granite models are trained on curated datasets drawn from the Internet, academic publications, code repositories, and legal and financial documents. The series encompasses a range of parameter sizes, from 3 billion to 34 billion, catering to different deployment needs.

## Foundation model architecture

A foundation model is an [AI](https://www.wikiprompt.org/wiki/artificial-intelligence) model trained on broad data at scale, enabling adaptation to a wide range of downstream tasks. IBM Granite models are decoder-only [transformers](https://www.wikiprompt.org/wiki/transformer), a common architecture for generative language models that predicts the next token in a sequence. The first foundation models, Granite.13b.instruct and Granite.13b.chat, each had 13 billion parameters, which was smaller than many contemporaneous large models. Later variants expanded the range from 3 billion to 34 billion parameters, offering flexibility for different computational budgets and use cases.

The training data curation emphasized quality and diversity, incorporating sources like academic papers, code datasets, and specialized domains such as legal and finance. This approach aimed to improve performance on enterprise-oriented tasks, a key focus for IBM's commercial offerings.

## Release and open-source code models

On May 6, 2024, IBM released the source code of four variations of Granite Code Models under the Apache 2.0 license, a permissive open-source license that permits free use, modification, and sharing. These models were made available on Hugging Face, a popular platform for hosting AI models. According to IBM's internal reports, the Granite 8b model outperformed [Llama 3](https://www.wikiprompt.org/wiki/meta-ai) on several coding-related benchmarks within a similar parameter range, highlighting its competitive performance in code generation and understanding.

The open-source release was part of a broader strategy to foster community adoption and collaboration, contrasting with some proprietary models from other vendors. IBM positioned Granite as a flexible alternative for enterprises seeking customizable AI solutions.

## Applications and ecosystem

Granite models are integrated into IBM's Watsonx platform, which provides tools for data analysis, generative AI, and model deployment. The models are used for tasks such as code generation, document summarization, and question answering, particularly in regulated industries like finance and law where data privacy and compliance are critical. By offering both open-source and cloud-based access, IBM aimed to serve a wide range of customers, from startups to large enterprises.

The release of open-source code models also enabled third-party developers to fine-tune and deploy Granite on their own infrastructure, including on [Amazon Web Services](https://www.wikiprompt.org/wiki/amazon-web-services) and other cloud providers. This ecosystem approach mirrored trends in the broader AI community, where open models like [LLaMA](https://www.wikiprompt.org/wiki/llama) and Mistral AI have gained traction.

## Performance and comparisons

IBM has published benchmarks showing Granite models achieving competitive results on standard NLP and coding tasks. For example, the Granite 8b model reportedly surpassed Llama 3 on coding benchmarks, despite having a similar number of parameters. However, independent verification of these claims is limited, and performance can vary depending on the specific task and evaluation methodology. The models are designed to be efficient, with smaller variants suitable for edge deployment and larger ones for cloud-based workloads.

## See also

- [Mistral AI](https://www.wikiprompt.org/wiki/mistral-ai), a company that also provides open-source models
- GPT
- [LLaMA](https://www.wikiprompt.org/wiki/llama)
- [Gemini](https://www.wikiprompt.org/wiki/gemini)

## References

(No external links are included per the article's style.)

## External links

- [Wikipedia: IBM Granite](https://en.wikipedia.org/wiki/IBM_Granite)

---
Source: https://www.wikiprompt.org/wiki/ibm-granite
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-10-10T06:05:27.083831+00:00