Introducing the IBM Granite 4.1 family of models

2026-04-29 · Source: IBM Research · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, medium

Summary

IBM has released the Granite 4.1 collection, an updated family of models designed for enterprise AI workflows. This release includes new small language models (SLMs) in 3B, 8B, and 30B parameter sizes, along with updated Granite speech, vision, embeddings, and Guardian models. The language models, trained on approximately 15 trillion tokens and extended to 512K context length, demonstrate strong performance in instruction following and tool calling, often matching or exceeding larger Granite 4.0 models. Granite Vision 4.1 is a vision-language model excelling in document understanding, particularly table, chart, and key-value pair extraction. Granite Speech 4.1 introduces multilingual recognition with a 5.33% word-error rate on the OpenASR Leaderboard and high-throughput NAR variants. Granite Guardian 4.1, fine-tuned on Granite 4.1 8B, enhances harm detection with expanded risk definitions. Finally, Granite Embedding Multilingual R2 supports over 200 languages for efficient semantic search. All models are Apache 2.0 licensed and optimized for enterprise deployment.

Key takeaway

For AI Engineers and MLOps teams building enterprise applications, the IBM Granite 4.1 collection offers a compelling suite of production-ready models. You should consider these Apache 2.0 licensed models for tasks requiring efficient instruction following, accurate document understanding, multilingual speech recognition, or robust harm detection. Their optimization for open-source inference runtimes like vLLM and SGLang allows flexible deployment, potentially reducing operational costs and improving reliability in your AI workflows.

Key insights

IBM's Granite 4.1 models offer modular, efficient, and governable AI solutions for enterprise workflows, prioritizing data quality over raw quantity.

Principles

• Enterprise AI benefits from combining diverse, specialized models.
• Data quality and staged refinement improve model performance.
• Efficiency and reliability are critical for production-ready enterprise AI.

Method

IBM's training philosophy for Granite 4.1 involves multi-phase training, starting broad and progressively annealing towards high-quality technical data, followed by multi-stage RL for distinct capabilities.

In practice

• Use Granite 4.1 SLMs for instruction following and tool calling tasks.
• Deploy Granite Vision 4.1 for document understanding, tables, and charts.
• Integrate Granite Guardian 4.1 for robust harm detection in AI pipelines.

Topics

• IBM Granite 4.1
• Small Language Models
• Enterprise AI
• Document Understanding
• Speech Recognition
• AI Safety Guardrails
• Multilingual Embeddings

Best for: CTO, VP of Engineering/Data, AI Architect, AI Engineer, MLOps Engineer, Director of AI/ML

Related on AIssential

• IBM Granite 3.0 models — IBM's Granite 3.0 models, including dense and MoE variants, are now available via Ollama under Apache 2.0.
• Granite 4.1, IBM Bob & building a quantum ecosystem — Enterprise AI is shifting from monolithic models to composable, specialized systems for cost-effective, sustainable operations.
• [P]Seeing models work is so satisfying — Larger datasets significantly improve model performance in speech-to-phone mapping tasks.
• What's New in Mellea 0.4.0 + Granite Libraries Release — Mellea 0.4.0 and Granite Libraries enable structured, verifiable, and safety-aware AI workflows by replacing probabilistic prompts with programmatic control.
• Mistral Small 4 in 8 mins! — Mistral Small 4 unifies reasoning, coding, and multimodal capabilities for enterprise applications under an Apache 2.0 license.
• Unlocking the Power of Language Models in Enterprise: A Deep Dive with Chris Van Pelt — Enterprise adoption of LLMs like GPT-3.5 and GPT-4 is a key focus.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by IBM Research.