Granite-4.0-H-Small is a 32B-parameter (32.2B) dense transformer from IBM, instruction-tuned through supervised finetuning, reinforcement learning alignment, and model merging over permissive open-source and synthetic datasets. It is published as an ungated, Apache 2.0 licensed release on Hugging Face, making the full weights openly available for self-hosting or downstream fine-tuning. The model's training mix covers twelve languages including English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese, giving it broad multilingual reach for global enterprise workloads.
Designed for general enterprise NLP, the model handles summarization, classification, question answering, retrieval-augmented generation, and code tasks, with structured chat support via system prompts and tool-calling through an OpenAI-style function schema for function integration and Fill-in-the-Middle code completion. Beyond Hugging Face, it is also offered as a deployable "IBM Granite 4.0 h-small" listing on AWS Marketplace, giving teams a managed deployment path alongside the open weights. Its combination of mid-sized parameter count, multilingual coverage, and open licensing makes it a practical fit for organizations that want a transparent, customizable foundation model for assistants, document pipelines, and tool-augmented workflows without depending on a closed API.
