Granite 13B
IBM's earlier enterprise foundation model, preceding the Granite 3 generation
Granite 13B is one of IBM’s first foundation models built for its watsonx enterprise AI platform, announced on September 7, 2023 alongside a family of Granite models aimed at business use cases rather than general chat. It has 13 billion parameters, uses multi-query attention with learned position embeddings, and was trained with the tokenizer from EleutherAI’s GPT-NeoX-20B, giving it an 8,192 token context window. IBM built it for tasks like document summarization, question answering, and text classification inside regulated industries such as banking and insurance, and it came with IBM’s standard indemnification for customers worried about training-data copyright exposure, a selling point competitors like OpenAI and Anthropic did not offer at the time. Granite 13B was proprietary and accessible only through watsonx, distinguishing it from IBM’s later Granite 3 generation, which shifted toward openly licensed weights on Hugging Face. It has since been superseded by newer Granite versions as IBM’s enterprise AI lineup moved toward open weights and broader benchmark parity with mainstream models.