IBM releases Granite 4.1: a full enterprise AI stack covering 3B/8B/30B language models, vision, speech, embeddings, and Guardian safety models under Apache 2.0.
Key Takeaways
Granite 4.1 8B instruct matches or outperforms Granite 4.0 32B MoE on instruction following and tool calling, at lower cost and simpler architecture.
Language models trained on ~15T tokens with multi-stage RL targeting distinct capabilities: instruction adherence, conversation quality, factual accuracy, math reasoning.
Context window extends to 512K tokens with no reported performance degradation on shorter tasks.
Granite Speech 4.1 2B hits 5.33% WER on OpenASR Leaderboard; a non-autoregressive NAR variant generates full sequences at once for higher GPU throughput.
Granite Embedding Multilingual R2 supports 200+ languages; Guardian 4.1 adds expanded risk definitions for agentic and safety use cases in pipelines.
