IBM Releases Granite 4.1 Model Family Across Language, Vision, Speech, and Embedding

IBM releases its Granite 4.1 model family, the company's most expansive AI model collection spanning language models in 3B, 8B, and 30B parameter sizes, plus vision, speech, embedding, and guardian capabilities. The 8B instruct model matches performance of the previous 32B Mixture-of-Experts model while using simpler architecture, trained on approximately 15 trillion tokens with support for up to 512K token context lengths. Notable releases include Granite Speech 4.1 2B achieving 5.33% word-error rate on OpenASR Leaderboard, Granite Vision 4.1 for document understanding, and Granite Embedding supporting over 200 languages. All models are released under Apache 2.0 license and available on watsonx and Hugging Face, optimized for open-source inference runtimes, with Granite Guardian 4.1 providing safety evaluation for any language model.

Granite 4.1 Language Models: Training Philosophy and Performance

• ▪The Granite 4.1 8B instruct model consistently matches or outperforms the Granite 4.0 32B Mixture-of-Experts model while using a simpler architecture
• ▪Granite 4.1 language models are trained on approximately 15 trillion tokens across multiple phases
• ▪Granite 4.1 language models support context lengths of up to 512K tokens
• ▪Granite 4.1 language models are offered in 3B, 8B, and 30B parameter base and instruct model sizes
• ▪IBM released the Granite 4.1 collection of AI models covering small language models, speech, vision, embeddings, and Guardian models

Multimodal Capabilities: Vision, Speech, and Embedding Models

• ▪Granite Speech 4.1 includes three variants: 2B, 2B Plus with richer transcription features, and 2B NAR with substantially higher throughput
• ▪Granite Vision 4.1 is a vision-language model designed for document understanding tasks including tables, charts, and key-value pair extraction
• ▪Granite Speech 4.1 2B achieves a 5.33% word-error rate, placing it among the top models on the OpenASR Leaderboard
• ▪Granite Vision 4.1 uses a feature injection scheme inspired by DeepStack that distributes visual information across multiple LLM layers
• ▪Granite Embedding includes a 97M-parameter embedding model that delivers state-of-the-art retrieval performance under tight resource constraints
• ▪Granite Embedding Multilingual R2 scales retrieval support to more than 200 languages while dramatically increasing context length
• ▪Granite Speech 4.1 2B NAR generates entire sequences at once rather than one token at a time, resulting in considerably better GPU utilization
• ▪ChartNet was created using a code-guided augmentation methodology and has been used for training Granite Vision 4.1

Enterprise AI Safety and Deployment Through Guardian and Open Licensing

• ▪Granite Guardian 4.1 is designed to evaluate LLM inputs and outputs for safety, quality, and correctness
• ▪Granite Guardian 4.1 was fine-tuned on top of Granite 4.1 8B and is a direct replacement for Granite Guardian 3.3 8B
• ▪Granite 4.1 models are available on watsonx and Hugging Face
• ▪All Granite 4.1 models are released under an Apache 2.0 license

Perspective of IBM Research

• ▪IBM released all Granite 4.1 models under Apache 2.0 license to enable broad commercial adoption

Perspective of Enterprise AI adopters

• ▪Organizations need safety guardrails like Granite Guardian 4.1 to evaluate LLM inputs and outputs before production deployment

Perspective of Open-source AI community

• ▪Granite 4.1's availability on Hugging Face enables researchers to access and experiment with the models through familiar infrastructure