IBM Granite 4.1: full family under Apache 2.0 with the 8B matching previous 32B performance.

The release includes LLMs at 3B/8B/30B, speech models (2B for ASR and translation), plus vision and embeddings. All at 512K context, trained on ~15 trillion tokens. The 8B delivering 32B-class performance at a fraction of the compute cost is the real story here. For cost-sensitive deployments, an Apache-licensed 8B that punches above its weight class is worth benchmarking.