IBM has launched the Granite 4.0 series of models, featuring four new versions: Micro (3B parameters, dense architecture), Micro H (3B parameters, hybrid architecture), Tiny H (7B total / 1B active parameters, hybrid MoE architecture), and Small H (32B total / 9B active parameters, hybrid MoE architecture). All models are released under the Apache 2.0 open-source license and are focused on tool calling and Retrieval-Augmented Generation (RAG) tasks.


These models require LM Studio 0.3.28 or a later version to run. Micro and Micro H are suitable for resource-constrained environments, while Tiny H and Small H balance performance and efficiency through their hybrid expert architecture.

Some users have asked for clarification on what 'hybrid architecture' means. The hybrid design combines the strengths of dense and MoE components, enhancing the model's capabilities while maintaining lower computational costs. For instance, Tiny H, with only 1B active parameters, can handle tasks at the level of a 7B model.

Regarding the models' actual performance, comprehensive benchmark test results have not yet been published. However, based on their design objectives, these models are expected to have a clear advantage in scenarios involving tool integration and document retrieval. Some developers have mentioned they have already begun testing, and the results are worth watching.

With its long-standing presence in the AI field, this release from IBM shows that the company is seeking a new position within the open-source ecosystem. The Apache 2.0 license for Granite 4.0 lowers the barrier for commercial use, potentially attracting more enterprises to try it.

The model download page currently does not provide file size information, which is not very friendly for mobile users. It is hoped that these practical details will be added later.

Reference link: https://lmstudio.ai/ibm/granite-4-micro