IBM Unveils Granite 4.1 Model Family

IBM unveiled Granite 4.1, updating its language, speech, vision, embedding, and guardrail models all at once.

• The language models come in 3B/8B/30B base and instruct versions, with improved performance over Granite 4.0.
• In particular, the 8B instruct showed results comparable to or better than Granite 4.0's 32B MoE, remaining competitive in instruction following and tool calling even without a reasoning variant.
• Training was conducted on roughly 15 trillion tokens, with later stages refined using technical, scientific, and math data, supporting up to 512K tokens of context.
• Multi-stage RL separately optimized instruction following, conversation quality, factuality, and math reasoning, reducing the side effects of single-stage tuning.

The multimodal and safety layers were also strengthened.

• Granite Vision 4.1 is a VLM for extracting tables, charts, and KVPs, with IBM touting stronger document understanding performance than comparable models.
• Granite Speech 4.1 2B recorded a 5.33% WER, and its Plus/NAR variants expand the choice between transcription quality and throughput.
• Granite Guardian 4.1 was updated as a moderation model that detects bias, toxicity, hallucination, agentic risk, and more, serving as the successor to Guardian 3.3 8B.
• Granite Embedding Multilingual R2 supports 200+ languages and longer context, tailored for large-scale multilingual search.

All models were released under Apache 2.0, with IBM presenting them as modular components for enterprise AI.