granite-4.2-3b

granite-4.2-3b

Reasons step by step before it answers, with tool calling, code generation and multilingual chat, in a compact 3B model from IBM. Supports a 128K context natively and is open under Apache 2.0.

Made by ibm-granite on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Granite-4.2-3B is a compact reasoning model from IBM's Granite Team. It is a decoder-only dense transformer built from Granite-4.1-3B-Base, and it thinks through a problem in think tags before it gives the final answer. You can run it in full thinking, non-thinking or low-effort mode, which lets you trade depth for speed on each query.

Who it suits: It suits developers and teams who want reasoning, code help, tool calling or multilingual chat from a model small enough for modest hardware.

Industry | Who uses it | How they use it | Example
Software development | Developers adding code help | Generate and reason about code | Ask for a function and pick low-effort mode when speed matters
Agent building | Teams that wire models to tools | Let the model reason about which tool to call | Send a task with tool definitions and read the call it makes
Multilingual support | Teams serving users in many languages | Hold a dialog in a language other than English | Chat in Japanese or Arabic, two of the tested languages

If you need a larger model, the family also has 8B and 30B dense sizes.

Where to use it:

• Python projects that use the transformers library
• Agent and tool-calling pipelines
• Long-document and multi-turn chat setups
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Math and code reasoning | Developers | A step-by-step solution, then a final answer
Tool calling | Agent builders | Function calls chosen after the model reasons about them
Long documents | Analysts | Answers over inputs up to the long context window
Multilingual dialog | Support teams | Replies in the user's language

How to set it up: What you need first:

• Python and the Hugging Face command line
• About 7.3 GB of memory for BF16 weights, or about 1.8 GB at 4-bit, which is an estimate
• The transformers library

1. Install the Hugging Face tools with pip.
2. Download the model files to your computer.
3. Load them with transformers and chat with the model.

pip install -U huggingface_hub
hf download ibm-granite/granite-4.2-3b

How to use it: Send the model chat messages, and choose a thinking mode per query. Full thinking is the default, non-thinking answers directly, and low-effort sits between them. The card also says the model reasons about which tools to call and why.

Example: You ask a multi-step math word problem. In full thinking mode the model works through the steps inside think tags and then states the answer. In non-thinking mode it answers directly, which is quicker but less deep.

For newcomers and experts: A good first try is one math or logic question in the default mode, then the same question in non-thinking mode to see the difference in speed and depth.

For experts, the design is a dense transformer with grouped query attention, using 40 query heads and 8 key-value heads. It uses rotary position embeddings with theta 10,000,000, a SwiGLU feed-forward block with hidden size 8192, RMSNorm with epsilon 1e-5 and untied input and output embeddings. The 3B size has an embedding size of 2560 and 40 layers.

You send a prompt and pick a thinking mode
The model reasons inside think tags
It decides whether a tool call is needed
You get the final answer or the call

Good to know:

Title: Layers by model size in the Granite 4.2 dense family
3B dense: 40 layers
8B dense: 40 layers
30B dense: 64 layers

• Context: it supports 128K natively, with a long-context extension to 512K.
• Languages: the card tested 12 languages, including English, Japanese and Arabic, and says others may work but are not fully tested.
• Best for: math and logic reasoning, writing code, calling tools, agent workflows and chat in many languages.
• License: Apache 2.0
• Cost to run: no API key or paid service named.
• Data it reaches: it downloads files from Hugging Face and then runs on your own machine.
• Not for: none named.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 3.7B
• Weights: about 7.3 GB in BF16; about 1.8 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 41,238
• Likes: 135
• License: apache-2.0
• Languages: en, de, es, fr, ja, pt, ar, cs
• Based on: ibm-granite/granite-4.1-3b-base
• Quantized versions: 67
• Fine-tunes: 14
• Tags: transformers, safetensors, granite, text-generation, granite-4.2, reasoning, thinking, tool-calling, conversational

Links

• Model: https://huggingface.co/ibm-granite/granite-4.2-3b
• Files: https://huggingface.co/ibm-granite/granite-4.2-3b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ibm-granite/granite-4.2-3b

Then follow the usage in its model card. Nothing in this kit runs by itself.