IBM Granite is a family of open-source artificial intelligence models developed by IBM, encompassing large language models (LLMs), vision-language models, speech models, embedding models, time-series forecasting models, and safety guardrail (Guardian) models. The Granite models are released under the permissive Apache 2.0 license and are designed primarily for enterprise use cases, emphasizing performance, efficiency, transparency, and trustworthiness.
Overview
IBM Granite models are developed by IBM Research and made available through IBM's watsonx platform, Hugging Face, Ollama, and other distribution channels. The models are built with a focus on enterprise workloads including retrieval-augmented generation (RAG), tool calling and agentic workflows, code generation, document understanding, and structured output generation.
Model Generations
Granite 3.0 (October 2024)
The third generation introduced dense decoder-only LLMs (8B and 2B parameters), mixture-of-experts (MoE) models (3B-A800M and 1B-A400M), Granite Guardian safety models, and a speculative decoder for accelerated inference. Granite 3.0 models were trained on over 12 trillion tokens across 12 natural languages and 116 programming languages.
Granite 3.1 (December 2024)
This release featured enhanced performance, expanded 128K context windows, new embedding models, and additional tools for LLM-based workflows.
Granite 3.2 (February 2025)
Granite 3.2 introduced experimental chain-of-thought reasoning capabilities (toggleable on/off), the first Granite Vision language model (Granite Vision 3.2 2B) focused on document understanding, new Guardian model sizes, time-series forecasting models (Tiny Time Mixers / TTM-R2.1) supporting daily and weekly predictions, and sparse embedding models.
Granite 4.0 (October 2025)
Granite 4.0 introduced a hybrid Mamba-2/transformer architecture with Mixture-of-Experts in select models. Key features include over 70% lower memory requirements and 2× faster inference compared to similar models, especially in long-context and multi-session scenarios. The models use no positional encoding (NoPE). Granite 4.0 was the first open model family to receive ISO 42001 certification and introduced cryptographic signing of model checkpoints. Available sizes include H-Small (32B total / 9B active), H-Tiny (7B total / 1B active), H-Micro (3B hybrid), Micro (3B traditional), and smaller variants down to 350M parameters.
Granite 4.1 (2026)
Granite 4.1 expanded the family to include language models (3B, 8B, and 30B parameters), speech models (2B with state-of-the-art transcription accuracy), vision models (4B for document understanding), Guardian models, and embedding models. The language models use a dense decoder-only architecture and are trained on approximately 15 trillion tokens with context lengths up to 512K tokens. They demonstrate competitive instruction-following and tool-calling performance without relying on long chains of thought.
Key Capabilities
- Language understanding and generation: Text generation, classification, summarization, entity extraction
- Tool calling and function calling: Integration with APIs and software tools for agentic workflows
- Code generation: Support for fill-in-the-middle (FIM) code completion across multiple programming languages
- Retrieval-augmented generation (RAG): Document-grounded question answering
- Structured JSON output: Native support for schema-constrained JSON generation
- Multimodal understanding: Document, chart, table, and image analysis (Vision models)
- Speech recognition: Multilingual transcription and translation (Speech models)
- Time-series forecasting: Zero-shot and fine-tuned forecasting (Tiny Time Mixers)
- Safety guardrails: Input/output risk detection, hallucination detection, and harm mitigation (Guardian models)
Trust and Transparency
IBM Granite models are built with several trust and transparency measures:
- Released under Apache 2.0 license
- Training data sources and methodologies disclosed
- ISO 42001 certification for AI management systems (Granite 4.0)
- Cryptographic signing of model checkpoints for provenance verification
- Top ranking on the Stanford Foundation Model Transparency Index
- Classified as a Class III Open Model on the Linux Foundation's Model Openness Framework
- Bug bounty program with HackerOne (up to $100,000)
- IBM provides uncapped indemnity for third-party IP claims against content generated by Granite models when used on IBM watsonx.ai
Availability
Granite models are available through IBM watsonx.ai, Hugging Face, Ollama, LM Studio, Replicate, OpenRouter, Docker Hub, NVIDIA NIM, Kaggle, and other platforms. They are supported in inference runtimes including vLLM, SGLang, llama.cpp, and MLX.
