IBM has officially launched Granite 4.0, its next-generation family of open-source language models, designed to deliver enterprise-ready performance with dramatically reduced hardware costs and memory requirements. Announced on October 2, 2025, Granite 4.0 introduces a hybrid Mamba/Transformer architecture that lowers GPU demands while maintaining competitive accuracy and speed.

The standout feature of Granite 4.0 is its breakthrough efficiency. Compared to conventional transformer-based large language models (LLMs), Granite 4.0 reduces RAM usage by up to 70% for long-context and multi-session workloads. This translates into substantial cost savings for enterprises, enabling deployment on more affordable GPUs without compromising throughput or reliability.

Granite 4.0 comes in multiple configurations to meet different needs:

• Granite-4.0-H-Small: 32B total parameters (9B active), ideal for customer support automation and multi-tool agent workflows.

• Granite-4.0-H-Tiny: 7B total parameters (1B active), optimized for edge, local, and latency-sensitive applications.

• Granite-4.0-H-Micro: A dense hybrid model with 3B parameters.

• Granite-4.0-Micro: A traditional 3B transformer for environments not yet ready for hybrid architectures.

All models are available in both Base and Instruct variants.

Open, Certified, and Secure

Granite 4.0 is open-sourced under the Apache 2.0 license and is the first open model family to achieve ISO/IEC 42001 certification—the international gold standard for AI governance, accountability, and transparency. To further ensure trust, IBM cryptographically signs all Granite 4.0 model checkpoints and has partnered with HackerOne for a bug bounty program.

Enterprise and Developer Access

Granite 4.0 models are available now on IBM watsonx.ai and partner platforms, including Hugging Face, Docker Hub, Kaggle, NVIDIA NIM, LM Studio, Dell Pro AI Studio, and more. Access via Amazon SageMaker JumpStart and Microsoft Azure AI Foundry will be available soon.

Benchmark Results

Granite-4.0-H-Small rivals much larger closed models on instruction-following and function-calling benchmarks, outperforming many peers at a fraction of the size. The models also shine on Stanford HELM and Berkeley Function Calling Leaderboard v3, proving competitive in complex enterprise workflows.

What’s Next

IBM plans to expand Granite 4.0 with additional model sizes and reasoning-enhanced variants before the end of 2025. Already, enterprises such as EY and Lockheed Martin have been testing Granite 4.0 at scale, highlighting its potential in real-world deployments.

A New Standard for Practical AI

With Granite 4.0, IBM is positioning itself at the forefront of enterprise AI by focusing on cost-effective, efficient, and trustworthy language models. For developers, enterprises, and researchers, Granite 4.0 offers a new path to build agentic AI workflows without the prohibitive costs traditionally associated with LLM deployment.