Last week at Tech Exchange, IBM was thrilled to release our third generation of Large Language Models, Granite 3.0.

The Granite 3.0 model series comprises of three types of models:

1. Workhorse models for enterprise tasks (granite-3.0-2b-instruct, granite-3.0-8b-instruct)
2. Efficient models for low latency and CPU-based deployments (granite-3.0-3b-a800m-instruct, granite-3.0-1b-a400m-instruct)
3. Guardrail models to support further trust and safety (granite-guardian-3.0-8b, granite-guardian-3.0-2b)

The Granite 3.0 models were trained on 5-6x the amount of data compared the first generations of the IBM Granite models, which corresponds to a significant lift in performance across general purpose and specialized tasks.

These models also represent the consolidation of multiple functionality into one model! Where previous versions of granite had separate models for English, multilingual, and code use cases, the Granite 3.0 model family consolidates all of these capabilities into one model (while expanding the multilingual support to 12 different languages).

Further these models were designed with Enterprise AI in mind, and demonstrate leading performance across a number of enterprise-focused benchmarks:

Finally - to support IBM's commitment to open source AI, all Granite 3.0 were released under an Apache 2.0 open source license.  Further, Granite goes beyond what almost any model provider does today, and is open and transparent about the training datasets used to create the Granite models.  In the Granite Technical Report (ibm.biz/granite-report), you can find a thorough accounting of all training datasets used.  IBM has also open sourced under Apache 2.0 its Data Prep Kit and data curation recipes that were used for all data engineering tasks that prepared the data for training.