News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
IBM Granite

IBM Granite

IBM's decoder-only foundation models, from 3 to 34 billion parameters

IBM Granite is a family of decoder-only foundation models built by IBM and first announced in September 2023. Trained on curated web, academic, code, legal and finance data, the series spans 3 to 34 billion parameters and includes code models released under an open licence.

Share on WhatsApp

  • Announced 7 September 2023, with a paper four days later
  • First models, Granite.13b.instruct and Granite.13b.chat, held 13 billion parameters
  • Family sizes span 3 billion to 34 billion parameters
  • Four code variants published on Hugging Face, 6 May 2024, under Apache 2.0
  • IBM reports Granite 8b ahead of Llama 3 on several coding tasks at similar size

What it is

IBM Granite is a line of decoder-only foundation models produced by IBM. A foundation model is one trained on very broad data so it can be adapted to many different jobs afterwards. The family was unveiled in September 2023 and its first members carried 13 billion parameters each, which was modest beside the largest systems of that period. Later additions range from 3 billion to 34 billion parameters. Training draws on hand-picked material from the open web, scholarly publishing, source code collections, plus legal and financial documents. IBM built the series mainly for its Watsonx cloud platform, and later published some code-focused variants openly.

How it came about

IBM announced the Granite line on 7 September 2023, with a supporting paper following four days later. The earliest entries were Granite.13b.instruct and Granite.13b.chat. On 6 May 2024 the company put four Granite Code Models on Hugging Face under Apache 2.0, a permissive open-source licence that lets anyone use, change and redistribute the software freely.

Inside the model

Every Granite model is a decoder-only transformer, the design that predicts text one token at a time. Sizes across the family run from 3 billion to 34 billion parameters, where a parameter is a learned weight inside the network. Training data was assembled from internet text, academic publications, code repositories and legal and finance documents. The first pair, Granite.13b.instruct and Granite.13b.chat, each held 13 billion parameters. IBM has not published details here on context length, token counts or training compute.

What it can do

Granite is built as a general foundation model that can be adapted to many downstream tasks. IBM's own evaluation found the 8 billion parameter Granite code model beating Llama 3 on several coding tasks among models of a comparable size. The 13 billion parameter originals were smaller than most leading systems of their day.

How to use it

Granite runs on IBM Watsonx and is also distributed through GitHub, Hugging Face and RHEL AI. In May 2024 IBM released four code model variants under Apache 2.0 and posted them on Hugging Face for anyone to download. Other models in the family remain proprietary.

Why it matters

By publishing code models under Apache 2.0, IBM joined the group of firms putting usable open weights into the open, alongside Mistral AI. The Granite family gave Watsonx a house-built option next to third-party models and added to the range of openly downloadable coding models.

Questions people ask about IBM Granite

What is IBM Granite?

A family of decoder-only foundation models from IBM, announced in September 2023. They are trained on curated web, academic, code, legal and finance data and are meant to be adapted to many different tasks.

How large are the models?

The first two, Granite.13b.instruct and Granite.13b.chat, each have 13 billion parameters. Later members of the family range from 3 billion up to 34 billion parameters.

Can I download them?

Yes for the code models. IBM put four Granite Code Models on Hugging Face in May 2024 under Apache 2.0, a permissive licence allowing free use, modification and sharing. Other Granite models stay proprietary.

Where do they run?

Granite was built for IBM's Watsonx cloud platform. The models are also available through GitHub, Hugging Face and RHEL AI.

Good reading and sources

Where this page comes from

The facts on this page come from the Wikipedia article "IBM Granite", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.

It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

The weekly digest

Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.