With the conversation around artificial intelligence shifting to weighty issues, such as return on investment and the cost of compute power, IBM Granite 3.0 models aim to boost ROI for businesses adopting AI by delivering efficiency, cost savings and safer AI usage, according to Sriram Raghavan (pictured), vice president of IBM Research AI at IBM Corp.

Since Granite 3.0 performs exceptionally well on tasks, such as retrieval-augmented generation, classification, summarization and entity extraction, these models enable companies to streamline their AI use cases and improve decision-making because they are purpose-built for business, Raghavan added. 

“The ROI conversation usually comes down to … what’s the most efficient and effective way that I can get to the outcome?” he said. “How do you help me get my AI working in that hybrid environment? Now with Granite 3.0, we started really with sort of fit-for-purpose models for some of our applications and software. They let you do everything enterprises want to do … RAG, summarization, classification, all of the normal fundamental building blocks of most enterprise use cases today.”

Raghavan spoke with theCUBE Research’s Dave Vellante at theCUBE on the Ground – IBM Analyst Forum event, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed how IBM Granite 3.0 models optimize ROI in the AI sector. (* Disclosure below.)

Exploring IBM Granite 3.0 models

Given that IBM Granite 3.0 activates only a subset of parameters during inference, they reduce computation time and cost, making it ideal for low-latency and on-device applications. As a result, these models ensure that AI outputs are closely aligned with unique business needs, Raghavan pointed out. 

“With Granite 3.0, we’re releasing two flagship language models, Granite 8 billion, Granite 2 billion,” he noted. “They get released in both a base version and an instruct version, but the point is that they are like the workhorse models. They’re being trained on over 12 trillion tokens. They cover both 12 languages, English, French, Spanish, Japanese that we double down on in IBM, and over 116 programming languages.”

To further optimize ROI, Granite 3.0 introduces new guardrails through the Granite Guardian models, which detect risks such as social bias, profanity and inaccuracies. This ensures safe AI interactions, especially for customer-facing applications, according to Raghavan. 

“We’re also releasing what we call Granite Guardian models, again 8 billion and 2  billion,” he said. “What do the Guardian models do? They in some respects give you sort of this cloak of protection around the main model. You have a model, you extend data and you get data out, but you always worry about, what if the model has toxicity, it has biased information? Somebody’s trying to jailbreak the model and get it to do what I don’t want it to do. With Granite Guardian models, you put them in front of the model and outside the model.”

Thanks to IBM Granite 3.0 models’ state-of-the-art performance, they are ideal for a myriad of use cases, such as customer care and consulting. This is based on the fine tuning of specific enterprise data, according to Raghavan.

“We have done some internal benchmarking of this latest generation of Granite models with carefully curated BPO benchmarks,” he explained. “The Granite models are doing phenomenally well because of the value we put to the data. That allows consulting to go in and offer a differentiated value proposition with the domain knowledge of running a BPO process.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE Research’s coverage of at theCUBE on the Ground – IBM Analyst Forum event: 

Watch the full event video below:

(* Disclosure: TheCUBE is a paid media partner for the IBM Analyst Forum event. Neither IBM Corp., the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE