RedHatAI/granite-3.3-8b-instruct

Granite-3.3-8B-Instruct: Enhanced Reasoning and Instruction Following

Granite-3.3-8B-Instruct is an 8-billion parameter language model from the Granite Team at IBM, released on April 16th, 2025. Built upon Granite-3.3-8B-Base, this model is fine-tuned for improved reasoning and instruction-following, featuring a 32K context length. A key differentiator is its support for structured reasoning using <think> and <response> tags, which clearly separate internal thought processes from final outputs.

Key Capabilities

• Enhanced Reasoning: Demonstrates significant gains on benchmarks like AlpacaEval-2.0 and Arena-Hard, with notable improvements in mathematics and coding tasks.
• Instruction Following: Excels in general instruction-following, making it suitable for AI assistants and business applications.
• Multilingual Support: Supports English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese.
• Long-Context Tasks: Capable of handling long document summarization and question-answering.
• Code-Related Tasks: Proficient in code generation and related tasks.

Performance Highlights

Granite-3.3-8B-Instruct shows strong performance across various benchmarks, particularly in math, with an 8.12 score on AIME24 and 69.02 on MATH-500. It also achieves 89.73 on HumanEval and 86.09 on HumanEval+, indicating robust coding capabilities. The model was trained on a balanced combination of permissively licensed data and curated synthetic tasks using IBM's Blue Vela supercomputing cluster.

Good for

• Developing AI assistants requiring strong instruction-following.
• Applications needing structured reasoning and clear output separation.
• Tasks involving mathematics, coding, and long-context understanding.
• Multilingual applications across the 12 supported languages.