IBM foundation models

In IBM watsonx.ai, you can use IBM foundation models that are built with integrity and designed for business.

The Granite family of IBM foundation models includes decoder-only models that can efficiently predict and generate language.

The models were built with trusted data that has the following characteristics:

  • Sourced from quality data sets in domains such as finance (SEC Filings), law (Free Law), technology (Stack Exchange), science (arXiv, DeepMind Mathematics), literature (Project Gutenberg (PG-19)), and more.
  • Compliant with rigorous IBM data clearance and governance standards.
  • Scrubbed of hate, abuse, and profanity, data duplication, and blocklisted URLs, among other things.

IBM is committed to building AI that is open, trusted, targeted, and empowering. For more information about contractual protections that are related to IBM indemnification, see the IBM Client Relationship Agreement.

The following foundation models from IBM are available in watsonx.ai:

For details about encoder models developed by IBM, see Supported encoder foundation models.

For details about third-party foundation models, see Third-party foundation models.

How to choose a model

To review factors that can help you to choose a model, such as supported tasks and languages, see Choosing a model and Foundation model benchmarks.

A deprecated foundation model is highlighted with a deprecated warning icon Warning icon. For details about model deprecation and withdrawal, see Foundation model lifecycle.

Foundation model details

The foundation models in watsonx.ai support a range of use cases for both natural languages and programming languages. To see the types of tasks that these models can do, review and try the sample prompts.

Granite 4 models

The Granite 4.0 foundation models belong to the IBM Granite family of models. The granite-4-h-small, granite-4-h-micro and granite-4-h-tiny are instruction-following models built for structured and long-context capabilities. The models use fine-tuning, reinforcement learning, and model merging to improve performance. Granite 4.0 offers better instruction handling and tool use, making it well-suited for enterprise tasks.

Usage
Designed to respond to general instructions and can be used to build AI assistants for multiple domains, including business applications. The model is capable of common generative tasks, including summarization, text classification, text extraction, question-answering, retrieval augmented generation (RAG), code related tasks, function-calling tasks, Fill-In-the-Middle (FIM) code and multilingual dialog use cases.
Size
  • Small: 30 billion parameters
  • Tiny: 7 billion paramters
  • Micro: 3 billion paramters
Token limits
Context window length (input + output): 131,072
Supported natural languages
English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese. Users may finetune Granite 4.0 models for languages beyond these languages.
Instruction tuning information
The Granite 4 models are fine tuned from Granite-4.0-H-Small-Base using a combination of open source instruction datasets with permissive license and internally collected synthetic datasets.
Model architecture
Decoder
Learn more
Read the following resources:

granite-3-1-8b-base

The Granite 3.1 8b foundation model is a base model that belongs to the IBM Granite family of models. The model extends the context length of Granite-3.0-8B-Base.

Usage

The Granite 3.1 base foundation model is a is a pre-trained autoregressive foundation model intended for tuning, summarization, text classification, extraction, question-answering, and other long-context tasks.

You can use the granite-3-1-8b-base foundation model for fine tuning purposes.

Size

8 billion parameters

Token limits

Context window length (input + output): 131,072

Supported natural languages

English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese. Users may finetune Granite 3.1 models for languages beyond these 12 languages.

Model architecture

Decoder

Learn more

Read the following resources:

Granite Instruct 3.3 Models

The Granite Instruct foundation models belong to the IBM Granite family of models. The granite-3-3-2b-instruct and granite-3-3-8b-instruct foundation models are Granite 3.3 Instruct foundation models. These models build on earlier iterations for improved reasoning, mathematics, coding, and instruction-following capabilities.

Usage
Designed to excel in long-context and instruction-following tasks such as summarization, problem-solving, text translation, reasoning, code tasks, function-calling, and more. Can be integrated into AI assistants across various domains.

Sizes - 8 billion parameters

Token limits

Context window length (input + output)

  • 8b: 131,072

Note: The maximum new tokens, which means the tokens generated by the foundation model per request, is limited to 16,384.

Supported natural languages

English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese. However, users may fine tune these Granite models for languages beyond these 12 languages.

Supported programming languages

The Granite Instruct models are trained with code written in 116 programming languages.

Instruction tuning information

The Granite Instruct models are fine tuned Granite Instruct base models trained on over 12 trillion tokens with a combination of permissively licensed open-source and proprietary instruction data.

Model architecture

Decoder

Learn more

Read the following resources:

granite-3-2-8b-instruct

Granite 3.2 Instruct is a long-context foundation model that is fine tuned for enhanced reasoning capabilities. The thinking capability is configurable, which means you can control when reasoning is applied.

Usage

Capable of common generative tasks, including code-related tasks, function-calling, and multilingual dialogs. Specializes in reasoning and long-context tasks such as summarizing long document or meeting transcripts and responding to questions with answers that are grounded in context provided from long documents.

Size

8 billion parameters

Token limits

Context window length (input + output): 131,072

Note: The maximum new tokens, which means the tokens generated by the foundation model per request, is limited to 16,384.

Supported natural languages

English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese

Instruction tuning information

Built on top of Granite-3.1-8B-Instruct, the model was trained using a mix of permissively licensed open-source datasets and internally generated synthetic data designed for reasoning tasks.

Model architecture

Decoder

Learn more

Read the following resources:

Granite Instruct 3.1 models

Usage
Granite Instruct foundation models are designed to excel in instruction-following tasks such as summarization, problem-solving, text translation, reasoning, code tasks, function-calling, and more.
Sizes
  • 2 billion parameters
  • 8 billion parameters
Supported natural languages
English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, Chinese (Simplified).
Supported programming languages
The Granite Instruct models are trained with code written in 116 programming languages.
Instruction tuning information
The Granite Instruct models are fine tuned Granite Instruct base models trained on over 12 trillion tokens with a combination of permissively licensed open-source and proprietary instruction data.
Model architecture
Decoder
Learn more
Read the following resources:

Granite Code models

Foundation models from the IBM Granite family. The Granite Code foundation models are instruction-following models fine-tuned using a combination of Git commits paired with human instructions and open-source synthetically generated code instruction datasets.

The granite-8b-code-instruct v2.0.0 foundation model can process larger prompts with an increased context window length.

Usage

The following Granite Code foundation models are designed to respond to coding-related instructions and can be used to build coding assistants:

  • granite-3b-code-instruct
  • granite-8b-code-instruct
  • granite-20b-code-instruct
  • granite-34b-code-instruct

The following Granite Code foundation models are instruction-tuned versions of the granite-20b-code-base foundation model that are designed for text-to-SQL generation tasks.

  • granite-20b-code-base-schema-linking
  • granite-20b-code-base-sql-gen
Sizes
  • 3 billion parameters
  • 8 billion parameters
  • 20 billion parameters
  • 34 billion parameters
Try it out

Experiment with samples:

Token limits

Context window length (input + output)

  • granite-3b-code-instruct : 128,000

  • granite-20b-code-instruct : 8,192

    The maximum new tokens, which means the tokens generated by the foundation model per request, is limited to 4,096.

  • granite-20b-code-base-schema-linking : 8,192

  • granite-20b-code-base-sql-gen : 8,192

  • granite-34b-code-instruct : 8,192

Supported natural languages

English

Supported programming languages

The Granite Code foundation models support 116 programming languages including Python, Javascript, Java, C++, Go, and Rust. For the full list, see IBM foundation models.

Instruction tuning information

These models were fine-tuned from Granite Code base models on a combination of permissively licensed instruction data to enhance instruction-following capabilities including logical reasoning and problem-solving skills.

Model architecture

Decoder

Learn more

Read the following resources:

Granite Guardian 3.1 models

The Granite Guardian foundation models belong to the IBM Granite family of models. The Granite Guardian foundation model are fine-tuned Granite Instruct models that is designed to detect risks in prompts and responses.

Usage
Granite Guardian foundation models are designed to detect harm-related risks within prompt text or model response (as guardrails) and can be used in retrieval-augmented generation use cases to assess context relevance (whether the retrieved context is relevant to the query), groundedness (whether the response is accurate and faithful to the provided context), and answer relevance (whether the response directly addresses the user's query).
Sizes
  • 8 billion parameters
Try it out
Experiment with samples:
Supported natural languages
English
Instruction tuning information
The Granite Guardian models are fine tuned Granite Instruct models trained on a combination of human annotated and synthetic data.
Model architecture
Decoder
Learn more
Read the following resources:

Granite time series models

Granite time series foundation models belong to the IBM Granite family of models. These models are compact, pretrained models for multivariate time series forecasting from IBM Research. The following versions are available to use for data forecasting in watsonx.ai:

  • granite-ttm-512-96-r2
  • granite-ttm-1024-96-r2
  • granite-ttm-1536-96-r2
Usage
You can apply one of these pretrained models on your target data to get an initial forecast without having to train the model on your data. When given a set of historic, timed data observations, the Granite time series foundation models can apply their understanding of dynamic systems to forecast future data values. These models work best with data points in minute or hour intervals and generate a forecast dataset with up to 96 data points per target channel.
Size
1 million parameters
Context length
Required minimum data points per channel in the API request:
  • granite-ttm-512-96-r2: 512
  • granite-ttm-1024-96-r2: 1,024
  • granite-ttm-1536-96-r2: 1,536
Supported natural languages
English
Instruction tuning information
The Granite time series models were trained on almost a billion samples of time series data from various domains, including electricity, traffic, manufacturing, and more.
Model architecture
Decoder
Learn more
Read the following resources:

Granite Vision 3.3 2b

The Granite Vision 3.3 2b is a compact and efficient vision-language foundation model that is built for enterprise use cases. The granite-vision-3-3-2b model introduces novel experimental features such as image segmentation, doctag generation, and multi-page support. The model also offers enhanced safety compared to earlier Granite vision models.

Usage
The granite-vision-3-3-2b foundation model is designed for visual document understanding, enabling automated content extraction from tables, charts, infographics, plots, diagrams, and more.
Size
2 billion parameters
Token limits
Context window length (input + output): 131,072
Supported natural languages
English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese.
Instruction tuning information
The granite-vision-3-3-2b foundation model was trained on a curated instruction-following dataset, comprising diverse public datasets and synthetic datasets tailored to support a wide range of document understanding and general image tasks. The model was trained by fine-tuning the granite-3-2b-instruct foundation model with both image and text modalities.
Model architecture
Decoder
Learn more
Read the following resources: