Granite-4.0-Micro: Compact 3B Model for Enterprise AI Workflows

Granite-4.0-Micro: Compact 3B Model for Enterprise AI Workflows

TL;DR

Granite-4.0-Micro is a 3B-parameter long-context instruction model from IBM, built for enterprise applications, assistants, and agentic workflows. It combines a 128K context length with instruction following, tool calling, retrieval-augmented generation, code tasks, and multilingual dialogue.

What the Model Is

Granite-4.0-Micro, now listed on AIOZ AI, is a long-context instruction model developed by the Granite Team at IBM.

Fine-tuned from Granite-4.0-Micro-Base, it supports instruction-driven business and technical workflows at a compact 3B-parameter scale.

It can handle tasks ranging from summarization and document question-answering to tool use, code completion, and multilingual dialogue.

How the Model Architecture Works

Granite-4.0-Micro uses a decoder-only dense transformer architecture with a 128K sequence length.

Its main architectural components include:

  • Grouped Query Attention (GQA): Uses grouped key-value heads for a more compact attention design
  • Rotary Positional Embeddings (RoPE): Encodes token positions across long sequences
  • SwiGLU: Supports the transformer’s feed-forward layers
  • RMSNorm: Normalizes hidden states throughout the network

Training and Alignment

The model was fine-tuned from Granite-4.0-Micro-Base using open-source instruction datasets with permissive licenses and internally collected synthetic data.

Its post-training process combines:

  • Supervised fine-tuning
  • Reinforcement learning alignment
  • Model merging
  • Structured chat formatting for instruction-driven interaction

Training and alignment were carried out on an NVIDIA GB200 NVL72 cluster, with the resulting instruct model tuned for instruction-following and tool-calling workflows.

Core Capabilities

  • Instruction following
  • Tool calling for agentic workflows
  • Retrieval-augmented generation
  • Summarization and information extraction
  • Code generation and Fill-In-the-Middle completion

Multilingual Support

Granite-4.0-Micro supports English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese.

This coverage makes it suitable for multilingual assistants and applications operating across multiple markets. Developers can also fine-tune Granite-4.0-Micro for languages beyond the supported list.

Inputs and Output

  • prompt: A question, request, instruction, or context provided to the model
  • output: The text response generated from the prompt

This interface supports direct instruction workflows as well as applications that add retrieval, tools, or agent orchestration around the model.

Key Technical Details

  • Technical identifier: Granite-4.0-Micro
  • Developer: Granite Team, IBM
  • Model type: Decoder-only dense transformer
  • Parameters: 3B
  • Maximum sequence length: 128K tokens
  • Attention: Grouped Query Attention (GQA)
  • Positional encoding: Rotary Positional Embeddings (RoPE)
  • Activation: SwiGLU
  • Normalization: RMSNorm
  • Base model: Granite-4.0-Micro-Base
  • License: Apache License 2.0

Download It on AIOZ AI

With 3B parameters, 128K context, tool calling, retrieval-augmented generation, code support, and multilingual dialogue, Granite-4.0-Micro brings broad enterprise AI capabilities into a compact model.

Download Granite-4.0-Micro on AIOZ AI and evaluate it for assistants, agents, and long-context workflows.

AIOZ AI - Granite-4.0-Micro
A 3B-parameter long-context instruct model from IBM, finetuned for enhanced instruction following and tool-calling. Supports 12 languages including English, Chinese, Arabic, and Japanese. Built on a dense Transformer with GQA, RoPE, SwiGLU, and 128K context length. Trained using SFT, RL alignment, and model merging techniques for enterprise applications.
AIOZ AI: The DePIN Ecosystem for AI

FAQ

Q1: What is Granite-4.0-Micro?

It is a 3B-parameter long-context instruction model from IBM, designed for enterprise applications, code tasks, and multilingual dialogue.

Q2: How much context does Granite-4.0-Micro support?

It supports a sequence length of up to 128K tokens.

Q3: What applications is the model suited for?

It is suited for business assistants, document question-answering, multilingual applications, agentic workflows, and long-context tasks.