IBM launches Granite 4.1 language models with improved performance and training methods

Apr 30, 2026

AI Summary

IBM has introduced Granite 4.1, a new series of open-source language models designed for enterprise applications. The models, available in sizes of 3B, 8B, and 30B parameters, demonstrate enhanced performance through a refined training process that emphasizes data quality and effective learning strategies.

Granite 4.1 includes three model sizes: 3B, 8B, and 30B, all licensed under Apache 2.0 for commercial use.
The 8B model outperforms the previous 32B MoE model in various benchmarks, indicating significant improvements in training techniques.
IBM utilized a multi-phase training approach, adjusting data mixtures and learning rates to enhance model performance.
A rigorous filtering system was implemented to ensure high-quality training data, with an LLM-as-Judge evaluating responses across multiple dimensions.
Reinforcement learning was applied in four stages to improve instruction-following capabilities and recover performance in specific areas like math.
The models are designed for predictable latency and reliable tool calling, making them suitable for production environments.
The 3B model is particularly noted for its efficiency in edge deployment scenarios, while the 30B model targets high-performance applications requiring advanced capabilities.
All models are available through IBM's API and on platforms like Hugging Face, with options for FP8 quantized variants to reduce memory usage.

ibmgranitemodelairesearch

IBM launches Granite 4.1 language models with improved performance and training methods

Related Stories

Thinking Machines Lab develops AI model for simultaneous conversation

ChatGPT Sees Increased Adoption Among Older Users in Early 2026

Optimizing Matrix Multiplication for Swift in LLM Training

arXivLabs Encourages Collaboration on New Features with a Focus on Privacy