IBM Launches Compact Granite 4.0 Nano Models for Mobile Devices

IBM expands its Granite 4.0 model family with new, smaller language models under the name Granite 4.0 Nano.

IBM announces the launch of Granite 4.0 Nano. These new, smaller language models are part of the recently launched Granite 4.0 models, a series of open-source language models. The new models are aimed at edge and on-device applications, with a focus on efficiency and low parameter count.

Four New Nano Models

The Granite 4.0 Nano series includes four language models: two models with hybrid architecture and two traditional transformer versions. The hybrid variants are Granite 4.0 H 1B (with approximately 1.5 billion parameters) and Granite 4.0 H 350M (with approximately 350 million parameters).

These models utilize a hybrid SSM-based architecture. Additionally, classic transformer alternatives are available with a similar parameter size, intended for compatibility with existing frameworks where hybrid architectures are still limited in support.

The models are optimized for use in edge environments and on devices, where computing power and memory are often limited. With this, IBM targets developers who need efficient models without sacrificing functionality.

Apache 2.0 License and Certification

Like previous Granite 4.0 models, these Nano variants are also released under an Apache 2.0 license. They are compatible with popular AI runtimes such as vLLM, llama.cpp, and MLX. The models are trained with over 15 trillion tokens and use the same training methodologies as the larger Granite models.

All Granite 4.0 Nano models carry the ISO 42001 certification for responsible model management. With this, IBM emphasizes the importance of transparency and responsible development within its AI offerings.

Compared to other small language models from Google, Alibaba, and LiquidAI, among others, Granite 4.0 Nano models, according to IBM, show strong performance on benchmarks for general knowledge, mathematics, programming, and safety. Also, on tasks important for agent-oriented workflows, such as instruction following and tool invocation, the models perform better than similar models in their weight class.

More information and technical details are available via Hugging Face, where the models can be freely accessed. IBM plans further expansions within the Granite 4.0 family in the future.