IBM Releases Granite 4.1 Model Family for Enterprise AI

IBM has introduced the Granite 4.1 family of models, its most extensive release to date, according to IBM Research. The collection includes new language, vision, speech, embedding, and guardian models designed for enterprise workloads, all available under the Apache 2.0 license.

At the core of the release are the Granite 4.1 language models, available in 3B, 8B, and 30B parameter sizes. These models deliver improved performance in instruction following and tool calling compared with previous Granite 4.0 versions. They are trained on around 15 trillion tokens and optimized for efficiency and reliability across enterprise applications.

Granite Vision 4.1 focuses on document understanding, including table, chart, and key-value pair extraction. It uses a training approach that combines real and synthetic enterprise data to enhance performance. IBM also released ChartNet, a large dataset created to strengthen chart comprehension.

Granite Speech 4.1 introduces multilingual speech recognition and translation models, offering options that balance throughput, latency, and transcription quality. One variant, Granite Speech 4.1 2B, achieves a 5.33 percent word error rate on the OpenASR Leaderboard. The models are designed for deployment in noisy or resource-constrained environments.

The Granite Guardian 4.1 model replaces version 3.3 and extends risk detection capabilities to evaluate safety, bias, and correctness in AI outputs. Meanwhile, Granite Embedding Multilingual R2 supports over 200 languages for efficient semantic search across large document sets.

All Granite 4.1 models can be accessed through platforms such as watsonx, Hugging Face, and Replicate, and are optimized for inference runtimes including vLLM, SGLang, and llama.cpp.