IBM launches new generation Granite language model
The vendor takes an open source approach with Granite 3.0. It trained the new model family with Nvidia H100 GPUs. Not all enterprises will be convinced to use the new AI models.
IBM on Monday released its new family of Granite language models under a fully permissive open source Apache 2.0 license.
The Granite 3.0 models include general-purpose language AI models such as Granite-3.0-8B-Instruct, Granite-3.0-2B-Instruct, Granite-3.0B-Base and Granite-3.0-2B-Base; guardrail and safety models such as Granite-Guardian-3.0-8B and Granite-Guardian-3.0-2B; and mixture-of-experts models including Guardian-3.0-3B-A800M-Instruct, Granite-3.0-1B-A400M-Instruct, Granite-3.03B-A800M-Base and Granite-3.0-1B-A400M Base.
The language models were trained on over 12 trillion tokens of data from 12 different languages and 116 programming languages, according to IBM. The 8B and 2B models will include support for extended 128K-context length and understand multi-modal documents by the end of the year.
The open source Granite Guardian 3.0 models enable developers to use safety guardrails by checking how an AI model responds to risks such as social bias, hate, violence and hacking. The models were also trained using Nvidia H100 GPUs, according to IBM.
Granite 3.0 models will support applications such as customer service, IT automation and cybersecurity.
The open source approach
The new Granite line comes as more vendors are gearing toward small language models and focusing on open source.
"Over the last 25 years, the gold standard for open source is an Apache license," IBM senior vice president and chief commercial officer Rob Thomas said during a media briefing about the new models. "We chose that for a very good reason."
IBM is betting that the future of AI is open, Constellation Research analyst Andy Thurai said.
"They offer smaller, more efficient, transparent models that are trained ethically and created responsibly to be the differentiator," Thurai said.
While IBM is not trying to make money by licensing the models, it wants organizations to use its Watsonx platform to run the models or fine-tune or build a new derivative model, Thurai added.
Compared to previous generations, the Granite 3.0 models appear to be more efficient and accurate, Moor Insights strategy analyst Patrick Moorhead said.
"This makes sense to me as the models weren't trained on 'world data,'" Moorhead said. World data includes the internet, entertainment and consumer video. Instead, IBM used enterprise data such as data from documents and spreadsheets.
Some weaknesses
Despite these advantages, IBM's weakness in this fast-moving generative AI market is not having "first mover" advantage, Thurai said.
"It will take a lot for IBM to convince customers to move to their platform," he said.
Moreover, while IBM has shown that the Granite models can be used for different enterprise use cases, customers will also need to see how the models perform with the natural language, Gartner analyst Arun Chandrasekaran said.
"They're definitely starting a bit on the weaker footing when it comes to applying large language models for like natural language processing-oriented use cases," Chandrasekaran said. These include summarization, content generation and question and answer.
IBM's core focus is currently code and code-related use cases, he said.
They offer smaller, more efficient, transparent models that are trained ethically and created responsibly to be the differentiator.
Andy ThuraiAnalyst, Constellation Research
The vendor must show how the model performs for consumer use cases such as creating recipes, making vacation plans, or getting the news, Moorhead said.
"Customers will need to be convinced that IBM can continue to invest and innovate," Moorhead said.
Customers also need to know when to use the Granite models versus a third-party model since both are available on its platform, Chandrasekaran said.
"IBM needs to provide more clarity, or even a selection tool, if you want to call it, that will enable customers to use the right model for the right use case by providing -- an evaluation tool that will enable developers to choose the right model and implement it very well," he said.
IBM also unveiled its next generation of Watsonx Code Assistant, powered by Granite code models. It offers general-purpose coding assistance across various code languages, including Java and Python.
The vendor also introduced the expansion to its AI-powered delivery platform, IBM Consulting Advantage. The platform contains AI agents, applications and frameworks to support IBM consultants with clients.
As part of the expansion, Granite 3.0 language models will become the default model in Consulting Advantage.
Esther Ajao is a TechTarget Editorial news writer and podcast host covering artificial intelligence software and systems.
