Home
Granite
Choose the right model, from sub-billion to 34B parameters, open-sourced under Apache 2.0.
Don’t sacrifice performance for cost. Granite outperforms comparable models1 across a variety of enterprise tasks.
Build responsible AI with a comprehensive set of risk and harm detection capabilities, transparency, and IP protection.
Base and instruction-tuned language models with new reasoning capabilities designed for agentic workflows, RAG, text summarization, text analytics and extraction, classification, and content generation.
Pre-trained model specialized on vision tasks for document and image understanding, supporting a range of file types and resolutions, and designed for efficient deployment in enterprise environments.
Decoder-only models designed for code generative tasks, including code generation, code explanation, and code editing, trained with code written in 116 programming languages.
Lightweight and pre-trained for time-series forecasting, optimized to run efficiently across a range of hardware configurations.
Safeguard AI with Granite Guardian, ensuring enterprise data security and mitigating risks across a variety of user prompts and LLM responses, with top performance in 15+ safety benchmarks.
NASA and IBM teamed up to create an AI Foundation Model for Earth Observations using large-scale satellite and remote sensing data.
Designed to significantly enhance understanding of user intent and increase the relevance of information and sources in response to a query.
Applying reasoning capabilities to Granite led to a significant jump in complex instruction following and maintained general performance and safety characteristics, whereas comparable models experienced degradation in these domains
Granite-3.2-8B-Instruct
DeepSeek-R1-Distill-Llama-8B
DeepSeek-R1-Distill-Qwen-7B
ArenaHard
55.23
| 17.17 |
10.36
Alpaca-Eval-2
61.16
21.85
15.35
IFEval
73.57
66.50
59.10
| MMLU |
66.93
45.80
50.72
| PopQA |
28.08
13.25
9.94
| TruthfulQA |
66.37
47.43
47.14
BigBenchHard
65.60
65.71
65.04
DROP
50.73
44.46
42.76
GSM8K
83.09
72.18
78.47
HumanEval
89.47
67.54
79.89
HumanEval+
86.88
62.91
78.43
AttaQ
85.99
42.87
42.45
*Prices shown are indicative, may vary by country, exclude any applicable taxes and duties, and are subject to product offering availability in a locale.
"At CrushBank, we've seen first-hand how IBM's open, efficient AI models deliver real value for enterprise AI – offering the right balance of performance, cost-effectiveness, and scalability. Granite 3.2 takes it further with new reasoning capabilities, and we're excited to explore them in building new agentic solutions."
David Tan
CTO
CrushBank
Use IBM Docling and open source Granite 3.1 to perform document visual question answering for various file types
Discover how to build an AI agent that can answer questions
In this tutorial, you will use the IBM® Granite-3.0-8B-Instruct model now available on watsonx.ai™ to perform custom function calling.
Quantize a pre-trained model in a few different ways to show the size of the models and compare how they perform on a task
Forecast the future based on learning with the TinyTimeMixer (TTM) Granite Model
Convert text into a structured representation and generate a semantically correct SQL query
Prompt tune a Granite model in Python using a synthetic dataset containing positive and negative customer reviews
View the full granite cookbook
Confidently deploy AI at scale with open source Granite models in production with Red Hat Enterprise Linux AI and watsonx. Build faster with capabilities like tool calling, 12 languages, chain of thought reasoning, and multi-modal adaptors
Download the excerpt to learn why IBM is recognized as a Strong Performer in “The Forrester Wave™: AI Foundation Models for Language, Q2 2024”
Read Signal65’s evaluation of Granite models in key areas for enterprise AI, including performance against competitors like Llama and Mistral.
Explore how Salesforce evaluates Granite against other LLMs for business applications in areas like accuracy, cost and speed.
The latest Granite models deliver new reasoning capabilities, a vision-supported model and increased efficiency – delivering competitive results at a lower cost”
In Mixture of Experts - episode 40, the panel tackles DeepSeek R1 misconceptions, explains model distillation, and dissects the open-source competition landscape.
Get curated selection of AI topics, trends and research sent directly to your inbox.
DeepSeek-R1 is a digital assistant that performs as well as OpenAI’s o1 on certain AI benchmarks for math and coding tasks, was trained with far fewer chips and is approximately 96% cheaper to use, according to the company.
IBM believes in the creation, deployment and utilization of AI models that advance innovation across the enterprise responsibly. IBM watsonx AI and data platform have an end-to-end process for building and testing foundation models and generative AI. For IBM-developed models, we search for and remove duplication, and we employ URL blocklists, filters for objectionable content and document quality, sentence splitting and tokenization techniques, all before model training.
During the data training process, we work to prevent misalignments in the model outputs and use supervised fine-tuning to enable better instruction following so that the model can be used to complete enterprise tasks via prompt engineering. We are continuing to develop the Granite models in several directions, including other modalities, industry-specific content and more data annotations for training, while also deploying regular, ongoing data protection safeguards for IBM developed models.
Given the rapidly changing generative AI technology landscape, our end-to-end processes are expected to continuously evolve and improve. As a testament to the rigor IBM puts into the development and testing of its foundation models, the company provides its standard contractual intellectual property indemnification for IBM-developed models, similar to those it provides for IBM hardware and software products.
Moreover, contrary to some other providers of large language models and consistent with the IBM standard approach on indemnification, IBM does not require its customers to indemnify IBM for a customer's use of IBM-developed models. Also, consistent with the IBM approach to its indemnification obligation, IBM does not cap its indemnification liability for the IBM-developed models.
The current watsonx models now under these protections include:
(1) Slate family of encoder-only models.
(2) Granite family of a decoder-only model.
* How smaller, industry-tailored AI models can offer greater benefits
https://www.ft.com/partnercontent/ibm/how-smaller-industry-tailored-ai-models-can-offer-greater-benefits.html
1Performance of Granite models conducted by IBM Research against leading open models across both academic and enterprise benchmarks - https://ibm.com/new/ibm-granite-3-0-open-state-of-the-art-enterprise-models