IBM unveils Granite 4.0, a fully open source, enterprise-ready LLM designed for high performance, efficiency, and U.S. leadership in AI against global competitors.
IBM has unveiled Granite 4.0, the latest generation of its homegrown open source large language models (LLMs), engineered to deliver high performance while significantly reducing memory and operational costs. Released under the Apache 2.0 license, cryptographically signed, and ISO 42001-certified, Granite 4.0 enables developers and enterprises to freely use, modify, and deploy the models for commercial purposes, reinforcing trust and transparency in open-source AI.
The launch positions the U.S. competitively against Chinese open source LLMs, such as Alibaba’s Qwen, and OpenAI’s gpt-oss family, symbolically reclaiming Western leadership in globally available open models. AI engineer Alexander Doria (aka Pierre-Carl Langlais) remarked, “ibm suiting up again after llama 4 fumbled,” adding, “we finally have western qwen.”
Granite 4.0 introduces a hybrid Mamba-transformer architecture, combining the contextual precision of transformers with the linear efficiency of Mamba for long documents. This design reduces GPU memory consumption by over 70%, enabling enterprise-scale inference on a single NVIDIA H100 GPU. Model variants include Granite-4.0-H-Small (32B parameters), H-Tiny (7B), H-Micro (3B hybrid), and Micro (3B transformer-only), offering scalability from edge to enterprise workloads.
Benchmarks show strong performance in instruction-following tasks, function calling, and retrieval-augmented generation (RAG), surpassing nearly all open-weight models while maintaining cost efficiency. Trained on 22 trillion tokens, the Granite 4.0 family is optimised for multi-agent workflows, customer support automation, and large-scale retrieval systems.
Available across Hugging Face, watsonx.ai, and partner platforms, Granite 4.0 supports enterprise pipelines while signalling a practical and safe open source AI alternative. With future expansions planned, including “Thinking” variants, Granite 4.0 exemplifies IBM’s strategy to deliver high-performance, transparent, and enterprise-ready open LLMs.
