IBM Granite 4.0 Nano Release Notes

IBM has released Granite 4.0 Nano, a series of small language models (SLMs) optimized for edge and on-device applications. These models aim to provide high performance within a minimal parameter footprint, reducing the need for massive parameter counts to achieve utility in specialized tasks.

Model Variants and Architectures

Granite 4.0 Nano consists of four instruct models and their corresponding base models, split across two architectural approaches to ensure compatibility across different runtimes:

  • Granite 4.0 H 1B: A dense LLM with approximately 1.5 billion parameters featuring a hybrid-SSM (State Space Model) based architecture.
  • Granite 4.0 H 350M: A dense LLM with approximately 350 million parameters featuring a hybrid-SSM based architecture.
  • Granite 4.0 1B and 350M: Traditional transformer-based versions of the 1B and 350M models, provided for workloads where hybrid architectures lack optimized support (such as llama.cpp).

Technical Specifications and Training

All Granite 4.0 Nano models were trained using the same improved methodologies, pipelines, and dataset of over 15 trillion tokens used for the original Granite 4.0 family.

Key deployment and governance standards include:

  • Licensing: Released under the Apache 2.0 license.
  • Runtime Support: Native architecture support for vLLM, llama.cpp, and MLX.
  • Certification: ISO 42001 certification for responsible model development, ensuring the models are built and governed to global standards.

Performance and Benchmarks

Granite 4.0 Nano models demonstrate increased capabilities compared to other models in the 0.2B to 2B parameter range, including those from Alibaba (Qwen), LiquidAI (LFM), and Google (Gemma).

General Capabilities

Across general benchmarks measuring Knowledge, Math, Code, and Safety, the Nano models show a significant increase in capabilities relative to their size.

Agentic Workflow Performance

The models outperform several similarly sized competitors on tasks critical for agentic workflows, specifically in:

  • Instruction Following: Measured by the IFEval benchmark.
  • Tool Calling: Measured by Berkeley's Function Calling Leaderboard v3 (BFCLv3).

Conclusion

Granite 4.0 Nano provides a flexible range of small-scale models that balance high performance with efficiency, offering both hybrid-SSM and traditional transformer architectures to maximize compatibility for on-device AI deployment.

Sources