IBM Releases Granite 4.1 Full Family — LLMs, Speech, Vision, Embeddings Under Apache 2.0

Hugging Face Blog·high signal

IBM released the complete Granite 4.1 family: dense decoder-only LLMs at 3B/8B/30B parameters (the 8B instruct matches the previous 32B MoE), two new speech models (Granite Speech 4.1 2B and 2B-NAR for autoregressive ASR with translation), plus vision and embedding models. All released under Apache 2.0 with 512K context length, trained on ~15 trillion tokens. The 8B model delivering 32B-class performance at a fraction of the compute cost is the headline for practitioners.