IBM Building Enterprise AI on a Granite Foundation

Throughout its storied history, IBM has always been a pioneer of enterprise solutions. This will continue as the world enters the AI era. At its annual TechXchange event, IBM announced the release of the Granite 3.0 family of generative AI models—its latest generation of AI models explicitly designed and trained for domain-specific enterprise AI functions.

This family is comprised of 11 models, ranging from general purpose and language to guardrails, safety and mixture-of-experts (MoE) that are available now for download through HuggingFace with an Apache 2.0 license. Granite 3.0 models are also integrated with IBMs watsonx platform and are the default models for powering the IBM Consulting Advantage AI-powered delivery platform.

Granite 3.0 is available in versions starting as low as one billion parameters, each coming in Base and Instruct variants. These variants are typical of many other open-sourced models, with the base variant not having any fine-tuning, allowing for developers to introduce their own. Meanwhile, the instruct variant has been fine-tuned on a broad set of IBM-curated instruction data for general tasks, general reasoning and math-based skills.

Two of the variants in the family, derived from the 2B and 8B general purpose versions, are Granite Guardian 3.0 2B and 8B. These variants further allow application developers to apply both passive guidance and active monitoring of user prompts and LLM responses on bias, fairness, hate language, abuse and profanity, to name a few. Guardian also provides checks specifically for RAG use cases, such as groundedness, context relevance and answer relevance.

Purpose-built for enterprise

While most other models are trained on public internet data, IBM has designed their latest models to be easy to train on enterprise data and governed for use in enterprise applications. This is a critical capability that is not only required for enterprise AI value creation but can be cost-prohibitive if not performed in an optimized, cost-efficient manner. To this end, IBM injects its models with enterprise data using its InstructLab method co-developed with its RedHat counterparts. According to IBM, in so doing, its Granite models are providing enterprise task-specific performance at smaller model sizes and is comparable in cost to other models, such as Mistral and Llama.

Additionally, enterprises typically have the need to use other open-sourced or proprietary models, which may result in the need for guardrail and safety requirements similar to the Granite 3.0 models. Consequently, Granite Guardian 3.0 can be used with open-sourced or proprietary models, as well.

A proof point that is even better than benchmarks

As part of its announcement, IBM provided a handful of benchmarks against other similar open-sourced LLMs demonstrating competitive performance and costs. However, what is perhaps an even stronger statement of IBMs belief in the power of these enterprise-class models is its own use of Granite 3.0 as foundational elements to both the IBM watsonx and IBM Consulting Advantage platforms.

The watsonx platform is IBM’s solution for helping enterprises develop their own AI-powered applications and assistants using their own data in a low-code and automated environment, while also delivering the accountability required with enterprise-grade AI solutions. At TechXchange, IBM announced the latest watsonx code assistant is powered by IBM Granite models.

IBM expects watsonx code assistant to accelerate the software development lifecycle, enhance developer productivity, and improve code quality across a wide range of programming languages, including—but not limited to—Java, Python, C, C++, Go, JavaScript, and Typescript.

Additionally, AI is evolving from its current form to what many are starting to call “agentic AI.” While currently, AI is primarily transactional and responsive in nature, agentic AI aims to introduce more autonomy, sophisticated reasoning and multi-step problem-solving through the use of AI agents. According to IBM, Granite 3.0 8B introduces into the watsonx platform key agentic AI capabilities, such as “advanced reasoning and a highly structured chat template and prompting style” for implementing complex tool workflows.

In its consulting business, the integration of Granite 3.0 with the IBM Consulting Advantage platform is further improving the productivity and efficiency of its 160,000 strong consultants with enhanced AI agents, applications and automation of repeatable IBM-developed consulting methodologies, best practices, and frameworks as they help IBM customers.

Building on a Granite foundation

Enterprise use cases will be pivotal as AI moves from experimentation to value-creation. A key factor is the ability of enterprises to easily adapt models to their own needs, use cases, workflows, and applications based on enterprise or proprietary data, policies, and processes.

Further, the evolution of AI to agentic AI will require even more advanced methods for efficiently training AI, and efficiently utilizing multiple models at the same time—whether in parallel or in series. Low code/no code environments and increasing levels of abstraction will also be critical as AI application development evolves from data scientists and AI-specific coding languages and frameworks to domain experts who might not be as deep in the science of AI, but are experts in the domains where AI seeks to create value.

To this end, IBM is contributing to the industry at large, a set of tools and models starting with its latest set of Granite 3.0 models and on through to the tools available with watsonx and IBM Consulting Advantage. Only in the fullness of time will we see if these foundational elements indeed help the future of AI come to fruition. However, it is clear that it is on these elements that IBM is building the future of its enterprise AI business.

source: eetimes.com