IBM released a series of Granite 4.2 models

Company IBM introduced an updated line of language models Granite 4.2, focused on corporate use and high accuracy in data processing tasks. The new series includes models with varying numbers of parameters, optimized for operation under limited computational resources and specific business scenarios.

The architecture of Granite 4.2 is based on an improved data preparation process, including filtering of low-quality content and deduplication at the document level. Developers applied the Instruction Tuning method to datasets consisting of 10 trillion tokens, which allowed for improved performance in logical inference and code writing tasks. The models support a context window of up to 128 thousand tokens.

Technical specifications include support for FP8 and INT4 quantization formats, which reduces video memory requirements when deployed on Nvidia H100 series GPUs. In tests on MMLU and HumanEval benchmarks, the models demonstrate an accuracy increase of 12-15% compared to the previous version 4.1. Model weights are available under the Apache 2.0 license, allowing their use in commercial products without restrictions.

The update is aimed at integration into the watsonx.ai platform, providing developers with tools for fine-tuning models on their own corporate data. Using Granite 4.2 allows for a reduction in inference time by 20% while maintaining response quality, which is critical for real-time automation support and financial report analysis systems.