unsloth/granite-3.3-2b-instruct

TEXT GENERATIONConcurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kPublished:Apr 18, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Granite-3.3-2B-Instruct is a 2-billion parameter language model developed by IBM, fine-tuned for enhanced reasoning and instruction-following across a 128K context length. It improves upon previous Granite versions in benchmarks like AlpacaEval-2.0 and Arena-Hard, with notable gains in mathematics and coding. The model supports structured reasoning using and tags, making it suitable for general instruction-following tasks and integration into AI assistants.

Loading preview...

Model Overview

Granite-3.3-2B-Instruct is a 2-billion parameter instruction-tuned language model developed by IBM, featuring a 128K context length. It is built upon Granite-3.3-2B-Base and demonstrates significant improvements in reasoning and instruction-following capabilities. The model supports structured reasoning through <think> and <response> tags, allowing for a clear separation between internal thought processes and final outputs.

Key Capabilities

  • Enhanced Reasoning: Improved performance on benchmarks measuring generic capabilities, mathematics, and coding.
  • Instruction Following: Designed for general instruction-following tasks.
  • Structured Output: Utilizes <think> and <response> tags for clear, structured reasoning.
  • Multilingual Support: Supports English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese.
  • Versatile Tasks: Capable of summarization, text classification, text extraction, question-answering, RAG, code-related tasks, function-calling, and long-context tasks.

Performance Highlights

Granite-3.3-2B-Instruct shows notable gains over its predecessors, Granite-3.1-2B-Instruct and Granite-3.2-2B-Instruct, particularly in:

  • AlpacaEval-2.0: Achieved 43.45 (compared to 27.17 and 34.51).
  • Arena-Hard: Scored 28.86 (compared to 23.3 and 24.86).
  • Math Benchmarks: Significantly improved on AIME24 (3.28) and MATH-500 (58.09).

Intended Use Cases

This model is designed for general instruction-following and can be integrated into AI assistants for various domains, including business applications.

Finetunes
2 models