Granite 4.1 8B
IBM · 2026-04-29 · 8.8B parameters
Granite 4.1 8B is IBM's 8B-class dense, decoder-only instruction model, fine-tuned from Granite 4.1 8B Base and supporting a 131,072-token context window. IBM says the Granite 4.1 language models were trained on approximately 15 trillion tokens and post-trained with supervised fine-tuning and a multistage reinforcement-learning pipeline to strengthen instruction following, chat, and tool use without relying on long chains of thought. IBM reports that the 8B instruct model matches or outperforms Granite 4.0 32B MoE, and releases its weights under Apache 2.0.