IBM’s work in generative AI has gone somewhat under the radar, but Thursday’s release of Granite 4.0 may change that.
Granite 4.0 distinguishes itself by employing an open-source hybrid Mamba/transformer architecture, which IBM says can run on lower-cost GPUs than comparable models. Additionally, Granite 4.0 models are the first in the world to be ISO 42001 certified and cryptographically signed, according to IBM.
Four small models are being released initially
The Granite family includes multiple generative AI models:
Granite-4.0-H-Small, a cost-effective mixture-of-experts model built on hybrid architecture.
Granite-4.0-H-Tiny, a 7-billion-parameter variant.
Granite-4.0-H-Micro, a 3-billion-parameter hybrid model.
Granite-4.0-Micro, a 3-billion-parameter conventional model.
Hybrid models use a combination of transformer models — or conventional large language model architectures — and Mamba architecture, which performs fewer calculations as the context increases.
Its primary advantage is improved inference efficiency. Granite 4.0 models require less RAM to run than conventional LLMs, IBM said, especially when responding to long queries or multi-session workloads.
As mentioned, Granite 4.0 models are ISO 42001 certified and cryptographically signed. The certification from the International Organization for Standardization aligns Granite 4.0 models with ISO’s AI management system within the context of an organization.
“Achieving ISO 42001 certification is a major milestone, not just for IBM Granite, but for the artificial intelligence and technology landscape,” said Avani Desai, CEO of Schellman, the certification body that assisted IBM with the certification. “As one of the first open-source AI model providers to be certified, IBM has now set an important precedent for how transparency, accountability, and innovation can coexist.”
IBM offers a bug bounty program via HackerOne for Granite, with payouts of up to $100,000.
More must-read AI coverage
More Granite 4.0 models are coming later this year
The Granite 4.0 family of models is available in IBM watsonx.ai, as well as from platform partners including HuggingFace, Kaggle, and Nvidia Nim. Availability on Amazon SageMaker JumpStart and Microsoft Azure AI Foundry is expected soon.
IBM solicited early feedback from EY and Lockheed Martin. This input and evaluations from the open-source community will be used to iterate on future versions of the models, the company said.
By the end of the year, IBM plans to release a ‘thinking’ version of Granite 4.0, optimized for more complex problems and additional model sizes, including Medium and Nano.
IBM first released Granite on Sept. 6, 2023, positioning it quickly as a business-focused solution. Because the models were open source, Granite’s primary competitors were Meta’s Llama models and the Qwen family of models.
Microsoft’s Copilot Pro subscription is now 365 Premium as Redmond rearranges how its AI products are packaged.