amdenterpriseai/aim-epyc-ibm-granite-granite-4-1-3b repository overview

⁠Granite 4.1 3B

Granite 4.1 3B is a 3 billion parameter long-context instruct model developed by IBM's Granite Team, featuring 131,072 token context length and enhanced post-training pipeline including supervised finetuning and reinforcement learning alignment. The model uses a decoder-only dense transformer architecture with 40 layers, Grouped-Query Attention (GQA), RoPE position embeddings, MLP with SwiGLU activation, and RMSNorm. Excels at summarization, text classification, text extraction, QA, Retrieval Augmented Generation (RAG), code tasks, function-calling, multilingual dialog, and Fill-In-the-Middle (FIM) code completions. Trained on open source instruction datasets with permissive licenses plus internally collected synthetic datasets, supporting 12 languages including English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese.