IBM Launches Granite Series Models, Entering the Competitive AI Field

2023-09-08

Global computing giant IBM is entering the competitive field of generative AI with the announcement of new generative AI capabilities and models for its previously announced Watsonx data science platform.

Named "Granite," the generative AI model is similar to popular LLM models like GPT-4, using input to analyze and summarize text content and output results. IBM stated in an official press release, "Just as granite is a durable and versatile material with many uses in construction and manufacturing, IBM believes these Granite models will bring lasting value to your business," but did not disclose further details about the models.

However, unlike other LLMs widely used among ordinary users, IBM positions its Granite models for enterprise and commercial clients, which is its strength. Developed by IBM Research, the Granite models - Granite.13b.instruct and Granite.13b.chat - employ a "decoder" architecture, similar to the ability of other large language models to predict the next word in a sequence.

The company further added, "At IBM, we focus on building business-oriented models. The Granite series models are no exception, so we trained them using various datasets - a total of 7TB before preprocessing and 2.4TB after preprocessing - to generate one trillion tokens, which are character sets with semantic meaning for the model."

Due to the enterprise focus of the Granite models, IBM stated that extra attention has been given to the security of these deployments. Each dataset used in training undergoes defined governance, risk, and compliance (GRC) review processes.

The initial Granite models are just the beginning: there are plans to release more models in other languages, and IBM is also preparing more models trained by IBM. IBM recently announced that it is now offering Meta's Llama 2-chat 70 billion parameter model for early access to specific customers and plans to make it widely available in late September. Additionally, IBM will host StarCoder, a large language model for code that includes over 80 programming languages, Git commits, GitHub issues, and Jupyter notebooks.

In addition to the new models, IBM will introduce new features in the watsonx.ai studio. Later this month, the company will launch the first version of the studio with prompt tuning, an efficient and cost-effective way for customers to train models on their own trusted data, adapting the base models to their unique downstream tasks. We will also introduce a synthetic data generator to help users create artificial tabular datasets from custom data patterns or internal datasets.

RECENT AI NEWS

RECENT AI TOOLS