Granite 4.0 Micro

Granite 4.0 Micro - Dense long-context instruct model

Granite 4.0 Micro is a 3B-parameter, decoder-only dense transformer instruct model from IBM's Granite 4 family designed for long-context and tool-enabled agentic workflows. The model is released under Apache 2.0 and is presented as an alternative when Mamba2-specific (hybrid) runtimes are not yet optimized.

Defining Characteristics

• 3B parameters and a dense (non-hybrid) decoder-only transformer architecture.
• Documented long-context capability with a 128K token sequence length.
• Instruction-finetuned with supervised finetuning and alignment (including RL-based techniques) for improved instruction following and tool-calling.

Core Design

• Uses attention with GQA, RoPE positional embeddings, SwiGLU MLPs, RMSNorm, and shared input/output embeddings as core components.
• Trained on a mixture of permissively licensed public datasets, internally generated synthetic data, and selected human-curated examples.

Tool & Coding Support

• Explicit tool-calling/function-calling workflow and a chat prompt template that formats tools and tool responses; examples follow OpenAI-style function schemas.
• Supports fill-in-the-middle (FIM) code completion with the documented FIM tags and provides examples for JSON-constrained output.