- IBM's third generation of Granite AI models include versions with 8 billion and 2 billion parameters
- The company has chosen to release these models under an Apache 2 license, an open-source approach to AI
- Analysts said the third generation of Granite models are an improvement over previous iterations
IBM has launched its third generation of Granite artificial intelligence (AI) models which power its watsonx platform. The company’s investment in improving foundation models and open-source licensing reflects a broader trend in the AI industry: a focus on creating value rather than just pushing the boundaries of model size.
"It’s all about how you create value now," said Rob Thomas, IBM’s SVP of software and chief commercial officer.
IBM's latest generation of models — called Granite 3.0 — comes with a few differentiating features, including open source licensing. IBM has chosen to release both of its Granite 3.0 models under an Apache 2 license. "The details of the license matter a lot," Dario Gil, IBM’s SVP and director of Research said. This approach allows businesses to modify and optimize models without restrictive licensing.
The decision to use the Apache 2 license is part of IBM’s broader strategy to enable wide collaboration when developing AI. "It's completely changing the notion of the how quickly businesses can adopt AI when you have a permissive license that enables contribution, enables community, and ultimately, enables wide distribution," Thomas said.
IBM’s commitment to open source "goes back over 20 years, starting with Linux," noted Nick Patience, VP and AI Practice Lead at The Futurum Group — and this approach could attract a broader base of developers and increase industry adoption. However, success will depend on the AI community's willingness to build on these models and whether IBM can balance an open ecosystem with profitability.
Picking up steam
According to Thomas, the company’s generative AI (GenAI) business has grown to $2 billion across technology and consulting.
“I'm not sure we've ever had a business that is scaled at this pace, which I think is a reflection of, one, the client interest in the topic, and two, our differentiated play around open source," Thomas noted.
He also highlighted IBM’s progress in foundation models, noting that few companies globally are capable of investing at the necessary scale.
The watsonx platform is IBM’s AI platform that offers tools for building, training and deploying AI models. Granite is now the default AI model family in watsonx, especially for IBM’s consulting services and AI assistants.
So far, Granite models have been most commonly used by enterprises for targeted tasks like customer service, IT automation and digital labor. "These are the main use cases where clients have shown the most interest so far," Thomas said.
The next generation of models, Granite 3.0, come in versions with 8 billion and 2 billion parameters. Parameters are the internal values a model adjusts during training to improve its predictions—more parameters generally lead to more sophisticated pattern recognition.
Compared to some of the largest models on the market, such as OpenAI’s GPT-4 or Google’s PaLM, which have hundreds of billions of parameters, IBM’s Granite models are relatively small. GPT-4 is reported to have up to 175 billion parameters, making IBM’s 8 billion-parameter model modest in size. Although, smaller models like Granite 3.0 can be more efficient, particularly in handling specific tasks such as data retrieval, summarization and classification, and they consume less computational power.
Plus, the Granite models still leverage 12 trillion tokens of training data, explained Gil. "We joke that these are small models, but they're not really small. A massive amount of data that has gone into it," he said, noting the models have been fine-tuned for enterprise tasks like summarization and entity extraction and trained across 12 languages and 116 programming languages.
AI at the edge
IBM is also differentiating through its efforts to optimize AI performance on smaller devices, like those found at the network edge, through its Mixture-of-Experts architecture.
These models allow AI to operate with low latency on small devices, a growing area of interest for industries looking to bring AI capabilities closer to where data is generated. "IBM has an advantage here," Patience said, noting the potential for low-latency applications in sectors requiring fast AI inference, such as logistics and industrial automation.
Yet, how these models will perform in real-world environments remains to be seen. "Beyond pure model performance, broader factors like governance and the development of future AI agents are still critical," Patience added.
The third generation of Granite models are an improvement over previous iterations, said Analyst Patrick Moorhead of Moorhead Insights and Strategy, both in accuracy and efficiency. "For enterprises, these are competitive if not superior models," he told Fierce Network, adding that with this iteration IBM has successfully addressed prior concerns about performance.
While the introduction of Granite 3.0 reflects IBM’s ambition to remain a key player in enterprise AI, the broader question is whether the models will deliver meaningful impact in real-world applications.
"I want to see enterprises having success with Granite models to be the most convinced," Moorhead said. “I think enterprises will watch IBM and their string of developments, and if they keep seeing innovations, they will lock-in.”