Granite 4.0 Micro
IBM · 2025-10-02 · 3.4B parameters
Granite 4.0 Micro is IBM's 3.4-billion-parameter, long-context instruct model, fine-tuned from Granite-4.0-Micro-Base and built as a decoder-only dense Transformer with GQA, RoPE, SwiGLU, and RMSNorm. IBM positions this conventional-attention variant as an alternative when Mamba2 support is not optimized (for example, llama.cpp and PEFT); its documented capabilities include instruction following, tool calling, RAG, code tasks, and multilingual dialogue. The weights are released under Apache 2.0 as part of the ISO 42001-certified Granite 4.0 family.