IBM Release Its New Granite 4.0 Hybrid AI Models with Lower Memory and Hardware Costs

Oct 3, 2025- by CMI| Category : Information and Communication Technology1 min read
IBM Release Its New Granite 4.0 Hybrid AI Models with Lower Memory and Hardware Costs

IBM released Granite 4.0, its new family of open large language models (LLMs), showcasing a hybrid Mamba/transformer architecture to reduce memory requirements and hardware costs in the machine learning industry.

IBM, Granite 4.0 models can run on significantly cheaper GPUs while maintaining performance, as stated by IBM.

The models are open‑source under the Apache 2.0 license and are the first open models to receive ISO 42001 certification, confirming their alignment with international standards for AI security, governance, and transparency. All Granite 4.0 checkpoints are cryptographically signed to verify provenance and authenticity.

Granite 4.0 is accessible via IBM watsonx.ai and distributed through platform partners such as Dell Technologies, Docker Hub, Hugging Face, Kaggle, LM Studio, NVIDIA NIM, Ollama, OPAQUE, and Replicate. Support for access via Amazon SageMaker JumpStart and Microsoft Azure AI Foundry is planned to follow.

The release includes Granite-4.0-H-Small (32B parameters, 9B active), Granite-4.0-H-Tiny (7B parameters, 1B active), and Granite-4.0-H-Micro (3B parameters). A conventional transformer variant, Granite-4.0-Micro, is also available for platforms that do not yet support hybrid architectures.

Benchmarking shows Granite 4.0 models outperform prior Granite releases. The Granite‑4.0‑H‑Small model outpaces virtually all open‑weight models on the Stanford IFEval benchmark, trailing only Llama 4 Maverick. On the Berkeley Function Calling Leaderboard v3, Granite 4.0 holds its ground against much larger models—delivering comparable performance at a lower cost.

Before launch, enterprise partners such as EY and Lockheed Martin tested Granite 4.0. IBM also teamed up with HackerOne to run a bug bounty program, offering up to $100,000 for vulnerabilities or jailbreak exploits.

Granite 4.0 was trained on a 22 trillion‑token, enterprise‑focused corpus. Instruction‑tuned models are available now, while reasoning‑optimized models are slated to release later this fall. IBM plans to roll out more variants by year-end—including Granite 4.0 Medium and Granite 4.0 Nano.

Share this story

About the author

Shivani Latey

is a talented content writer with over three years of experience specializing in crafting informative and engaging articles and blog posts. Known for her ability to simplify complex topics, she ensures that her content is clear, accessible, and resonates with a wide range of audiences. Her strengths lie in distilling intricate concepts into digestible, well-structured pieces that maintain reader interest and drive engagement. With expertise in digital marketing and SEO, she is adept at creating content that not only informs but also enhances brand visibility and online presence. she create content for vrious industries such as Cosmetic Ingredients, Polymer and Resins, and Green Chemicals.

Driven by a passion for clarity and precision, she consistently strives to produce high-quality content that stands out in a crowded digital landscape. Her exceptional writing skills are complemented by a strong attention to detail and an innate ability to conduct thorough research, ensuring that every piece she produces is both credible and impactful. Always focused on delivering value, she takes great pride in producing content that aligns with industry trends while offering fresh perspectives that keep readers engaged and informed.