IBM Granite 4.1 LLMs: 8B Dense Model Matches Prior 32B MoE with 512K Context and Apache 2.0 License

Summary

IBM released Granite 4.1, a family of open-source, dense decoder-only language models available under the Apache 2.0 license in three sizes: 3B, 8B, and 30B parameters. The headline technical achievement is efficiency: Granite 4.1-8B matches or surpasses the previous Granite 4.0-H-Small (a 32B mixture-of-experts model) despite being a simpler dense architecture with fewer active parameters. Context window was extended from 4K to 512K tokens through a five-phase pre-training pipeline totaling app