IBM Granite 4.1 Review: 10-Model Family, 512K Context, Apache 2.0, and the 8B-Beats-32B Story

IBM Granite 4.1 (April 29, 2026) is a ten-model family covering language (3B/8B/30B), vision, speech, safety, and embeddings — all Apache 2.0. The 8B language model outperforms IBM's own prior 32B MoE, hits 85.37% HumanEval and 68.27% BFCL v3 tool calling, and costs $0.05/$0.10 per million tokens on OpenRouter. IBM also ships uncapped IP indemnification for watsonx customers — rare in enterprise AI. Not a reasoning model. SimpleQA factual recall is poor. GPQA is mediocre. Strong for enterprise agentic workflows. Rating: 4/5.