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
Granite 3.0

Granite 3.0

IBM's third-generation open enterprise language models, from 2B to 8B parameters

Granite 3.0 is the third generation of IBM's Granite family of large language models, released on 21 October 2024. The collection mixes dense and mixture-of-experts models, guardrail models and a speculative decoder, all under Apache 2.0, with training data and methods disclosed in a technical paper.

Share on WhatsApp

  • Released 21 October 2024 as the third Granite generation
  • 8B Instruct trained on over 12 trillion tokens, 12 natural and 116 programming languages
  • All models under Apache 2.0, with training data and methods disclosed
  • Dense, MoE, guardrail and speculative decoder variants in one release
  • Trained on IBM's Blue Vela cluster with NVIDIA H100 GPUs

What it is

Granite 3.0 is the third generation of IBM's Granite series of large language models and companion tools. The headline model, Granite 3.0 8B Instruct, is an instruction-tuned dense decoder-only model of 8.1 billion parameters, trained on more than 12 trillion tokens spanning 12 natural languages and 116 programming languages. IBM positions it as a building block for enterprise workflows, tool use and retrieval augmented generation. The release also includes 2B dense models, two mixture-of-experts models for low latency, two guardrail models and a speculative decoder. Everything ships under the permissive Apache 2.0 licence.

How it came about

IBM released Granite 3.0 on 21 October 2024 as the third generation of the Granite line. Earlier generations focused on specialised domains such as finance, legal, code and academia; this one also targets general use. The team ran thousands of experiments on data mixtures, plus hundreds of small 1-2B explorations, to settle the final recipes. IBM says it published a technical paper detailing training datasets and methods. A note on the announcement page points to Granite 3.2, released 25 February 2025, adding experimental reasoning and document understanding.

Inside the model

The 8B Instruct model is a decoder-only dense transformer: 40 layers, embedding size 4096, 32 attention heads with 8 key-value heads (grouped-query attention), RoPE position embeddings, a SwiGLU MLP of hidden size 12800, RMSNorm and shared input/output embeddings. It was pretrained on 12 trillion tokens, then fine-tuned with supervised data, reinforcement learning alignment and model merging. The MoE variants use 32 or 40 experts with a top-k of 8, activating 400M or 800M parameters. Training ran on IBM's Blue Vela supercomputing cluster with NVIDIA H100 GPUs, powered by 100% renewable energy. IBM's Power learning-rate scheduler and the open-sourced Data Prep Kit supported the process.

What it can do

IBM says Granite 3.0 8B Instruct matches similarly sized open models from Meta and Mistral AI on Hugging Face's OpenLLM Leaderboard v2, and leads the evaluations on RAGBench across its 11 datasets for faithfulness and correctness. It performs strongly on IBM's proprietary and public cybersecurity benchmarks, and outperformed leading open models in its weight class across six tool-calling benchmarks including Berkeley's Function Calling Leaderboard. Intended uses include generation, classification, summarisation, entity extraction, customer service, code generation and editing, and agentic tool calling. Multilingual performance may lag English; few-shot examples can help.

How to use it

All Granite 3.0 models are released under Apache 2.0. The 8B and 2B Instruct models and both Guardian models were made available for commercial use on the IBM watsonx platform, and through partners including Google Vertex AI, Hugging Face, NVIDIA NIM microservices, Ollama and Replicate. Weights are downloadable from Hugging Face, with recipes and guides on GitHub and a Granite Playground for trying the 8B Instruct model.

Safety and testing

IBM pairs the release with Granite-Guardian-3.0-8B and 2B guardrail models for input and output screening. The company says the models were trained on curated enterprise data filtered for objectionable content, with attention to governance, risk, privacy and bias, and reports strong robustness on the AttaQ adversarial-prompt benchmark. IBM publishes a Responsible Use Guide and discloses pretraining datasets. It offers uncapped indemnity for third-party IP claims against its own models. The model card warns that inaccurate, biased or unsafe outputs can still occur.

Reception and debate

IBM claims the 8B Instruct model rivals similarly sized open models on academic benchmarks and leads on enterprise evaluations such as RAGBench and tool calling. The company stresses its Apache 2.0 release and training-data disclosure as a contrast to closed or restrictively licensed rivals.

Why it matters

Granite 3.0 continues IBM's line of openly licensed enterprise models and is designed to be fine-tuned cheaply for organisation-specific needs, including through InstructLab, IBM's open source approach using synthetic data and phased training. IBM frames smaller fit-for-purpose models like Granite as a route to frontier-level results at lower cost. The family was succeeded by Granite 3.2 in February 2025.

Questions people ask about Granite 3.0

What is Granite 3.0?

It is the third generation of IBM's Granite family of large language models, released on 21 October 2024. The collection covers dense models of 2B and 8B parameters, mixture-of-experts models, guardrail models and a speculative decoder, aimed at enterprise use.

How can I use it?

Weights are published under the permissive Apache 2.0 licence on Hugging Face, and the models run on IBM watsonx as well as Google Vertex AI, NVIDIA NIM microservices, Ollama and Replicate. IBM also provides a Granite Playground, GitHub recipes and InstructLab for fine-tuning.

What languages does it support?

The models cover 12 natural languages: English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch and Chinese, plus 116 programming languages. IBM notes multilingual performance may be weaker than English and suggests few-shot examples.

How large is the main model?

Granite 3.0 8B Instruct has 8.1 billion parameters across 40 layers, with a 4096-token sequence length at release. IBM planned to expand context windows to 128K tokens later in 2024.

What came next?

IBM released Granite 3.2 on 25 February 2025, adding experimental reasoning capabilities and visual understanding with a focus on documents.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Granite 3.0 yet, so the facts on this page come from the sources listed above: ibm.com/new/announcements/ibm-granite-3-0-open-state-of-the-art-enterp, huggingface.co/ibm-granite/granite-3.0-8b-instruct. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    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.