IBM Launches Granite 3.0 AI Models for Businesses

IBM (NYSE: IBM) has announced the launch of Granite 3.0, its most advanced family of AI models to date, during the annual TechXchange event. This third-generation suite of Granite flagship language models is designed to outperform or match similarly sized models from leading providers across various academic and industry benchmarks, emphasizing strong performance, transparency, and safety. 

In line with IBM’s commitment to open-source AI, Granite models are released under the permissive Apache 2.0 license, providing unique advantages in performance, flexibility, and autonomy for enterprise clients and the broader community. The Granite 3.0 family features several models, including Granite 3.0 8B Instruct, Granite 3.0 2B Instruct, Granite Guardian 3.0 8B, and various Mixture-of-Experts models designed for specific tasks. 

The new Granite 3.0 models are tailored for enterprise AI applications, excelling in tasks like Retrieval Augmented Generation (RAG), classification, summarization, and entity extraction. These compact models are versatile enough to be fine-tuned with enterprise data and seamlessly integrated across various business environments. 

While many large language models (LLMs) are trained on publicly available data, significant amounts of enterprise data remain underutilized. By combining smaller Granite models with proprietary data—especially through the innovative InstructLab technique introduced by IBM and RedHat—businesses can achieve competitive task-specific performance at a fraction of the cost, with savings ranging from 3x to 23x compared to larger frontier models in initial proofs-of-concept. 

Granite 3.0 reinforces IBM’s dedication to transparency, safety, and trust in AI products. The technical report and responsible use guide accompanying the release detail the datasets used for training, the filtering and curation processes, and comprehensive performance results against major benchmarks. IBM also provides an IP indemnity for all Granite models on watsonx.ai, boosting enterprise clients’ confidence in merging their data with these models. 

Granite 3.0 models showcase exceptional raw performance. For instance, on the OpenLLM Leaderboard, the Granite 3.0 8B Instruct model leads in overall performance against similar-sized open-source models from Meta and Mistral. Additionally, it excels in safety metrics across all dimensions, surpassing competitors on the AttaQ safety benchmark. On essential enterprise tasks like RAG and tool use, the Granite 3.0 8B Instruct model consistently outperforms similar models from Mistral and Meta. 

Trained on over 12 trillion tokens derived from 12 natural languages and 116 programming languages, the Granite 3.0 models employ a novel two-stage training method, leveraging extensive experimental results to optimize data quality and training parameters. By the end of the year, the 3.0 8B and 2B models are expected to support an extended 128K context window and multi-modal document understanding. 

IBM’s Granite Mixture of Experts (MoE) Architecture models, such as Granite 3.0 1B-A400M and Granite 3.0 3B-A800M, provide excellent performance-to-cost ratios, suitable for low-latency applications and CPU-based deployments. Additionally, IBM is releasing updated pre-trained Granite Time Series models that offer superior performance on major benchmarks, outperforming models ten times larger from Google and Alibaba. 

The introduction of Granite Guardian 3.0 marks a significant step in responsible AI, allowing developers to implement safety guardrails by evaluating user prompts and LLM responses for various risks. These models feature comprehensive risk and harm detection capabilities, including checks for social bias, toxicity, and context relevance. Extensive testing reveals that Granite Guardian 3.0 8B exceeds previous model generations in harm detection accuracy and matches specialized hallucination detection models. 

All Granite 3.0 models, along with the updated time series models, are available for download on Hugging Face under the Apache 2.0 license. The instruct variants of Granite 3.0 8B and 2B models and the Granite Guardian 3.0 8B and 3B models are also available for commercial use on IBM’s watsonx platform. Furthermore, selected Granite 3.0 models will be accessible as NVIDIA NIM microservices and integrated into Google Cloud’s Vertex AI Model Garden. 

IBM has expanded its open-source catalog of powerful LLMs through collaborations with partners such as AWS, Docker, Domo, Qualcomm, Salesforce, and SAP, integrating Granite models into their offerings to enhance enterprise choices globally. 

As IBM continues to advance enterprise AI, it offers a range of technologies—from models and assistants to tools for tuning and deployment tailored to unique business needs. The upcoming release of the next generation of watsonx Code Assistant, powered by Granite code models, will deliver general-purpose coding assistance across multiple programming languages, further enhancing enterprise capabilities. 

Additionally, IBM has announced a significant expansion of its AI-powered delivery platform, IBM Consulting Advantage. This platform empowers 160,000 consultants with AI agents, applications, and methods to deliver enhanced client value efficiently. The Granite 3.0 language models will serve as the default in Consulting Advantage, optimizing ROI for generative AI projects. 

IBM Consulting Advantage will also focus on cloud transformation and business operations, featuring domain-specific AI agents and applications that integrate IBM’s best practices to accelerate client transformations in various sectors, including finance, HR, and procurement.