ibm-granite/granite-3.2-8b-instruct

TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:Feb 17, 2025License:apache-2.0Architecture:Transformer0.1K Open Weights Cold

Granite-3.2-8B-Instruct is an 8-billion-parameter, long-context AI model developed by IBM, fine-tuned for enhanced reasoning capabilities. Built upon Granite-3.1-8B-Instruct, it leverages a mix of permissively licensed open-source and internally generated synthetic data for reasoning tasks. This model allows for controllable 'thinking' capabilities, making it suitable for general instruction-following and integration into AI assistants across various domains, including business applications.

Loading preview...

Overview

Granite-3.2-8B-Instruct is an 8-billion-parameter, long-context AI model developed by the IBM Granite Team, released on February 26th, 2025, under an Apache 2.0 license. It is an evolution of Granite-3.1-8B-Instruct, specifically fine-tuned to enhance reasoning capabilities. The model's training data combines permissively licensed open-source datasets with IBM's internally generated synthetic data, focusing on improving its ability to 'think' and process complex reasoning tasks. A key feature is the controllability of its thinking capability, allowing it to be activated only when required.

Key Capabilities

  • Enhanced Reasoning: Fine-tuned with synthetic data specifically for reasoning tasks, offering controllable 'thinking' functionality.
  • General Instruction Following: Designed to handle a wide array of instruction-following tasks.
  • Multilingual Support: Supports English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese, with potential for fine-tuning in other languages.
  • Long-Context Processing: Capable of handling long documents for summarization, question-answering, and other long-context tasks.
  • Diverse Applications: Excels in summarization, text classification, text extraction, question-answering, Retrieval Augmented Generation (RAG), code-related tasks, and function-calling.

Performance Highlights

Compared to its predecessor, Granite-3.1-8B-Instruct, and other 7-8B models like Llama-3.1-8B-Instruct and Qwen-2.5-7B-Instruct, Granite-3.2-8B-Instruct shows notable improvements in specific benchmarks:

  • Achieves 55.25 on ArenaHard and 61.19 on Alpaca-Eval-2, significantly outperforming its predecessor and other models in these areas.
  • Maintains strong performance in MMLU (66.79), GSM8K (81.65), HumanEval (89.35), and HumanEval+ (85.72).

Intended Use Cases

  • AI Assistants: Ideal for integration into AI assistants for general instruction-following.
  • Business Applications: Suitable for various business-oriented AI tasks.
  • Complex Problem Solving: Leveraging its enhanced reasoning, it can tackle problems requiring a structured thought process, as demonstrated by its ability to show detailed step-by-step reasoning when thinking=True is enabled during generation.