IBM has released the next generation models in its Granite family: Granite 3.2 8B Instruct, Granite 3.2 2B Instruct, Granite Vision 3.2 2B, Granite-Timeseries-TTM-R2.1, Granite-Embedding-30M-Sparse, and new model sizes for Granite Guardian 3.2.

Granite 3.2 8B Instruct and Granite 3.2 2B Instruct provide chain of thought reasoning that can be toggled on and off. According to IBM, chain of thought reasoning can be powerful, but requires significant computing power that isn’t needed for every task, which can lead to unnecessary usage. 

The company took steps to mitigate this by allowing this feature to be easily turned off when it’s not needed, and applying Thought Preference Optimization (TPO)-based reinforcement learning, which allows it to achieve greater performance on complex reasoning without compromising performance elsewhere, the company explained.

“The release of Granite 3.2 marks only the beginning of IBM’s explorations into reasoning capabilities for enterprise models. Much of our ongoing research aims to take advantage of the inherently longer, more robust thought process of Granite 3.2 for further model optimization,” IBM wrote in a blog post. 

Granite Vision 3.2B is a new multimodal model that was designed for document understanding tasks. According to IBM, this model matches or exceeds Llama 3.2 11B and Pixtral 12B on enterprise benchmarks including DocVQA, ChartQA, AI2D, and OCRBench. 

Granite-Timeseries-TTM-R2.1 extends the model’s forecasting capabilities to now offer daily and weekly predictions. Previously, it only supported forecasting for minutes and hours. 

Granite-Embedding-30M-Sparse is an evolution of the Granite Embedding models that now has the ability to learn sparse embeddings, in which their embedding size equals their vocabulary size, and can be significantly faster than dense embeddings for shorter text passages. 

The company is also releasing a 30% smaller Granite Guardian safety model, Granite Guardian 3.2 5B, that matches the performance of the previous generation. Granite Guardian also has a new feature, verbalized confidence, providing a “more nuanced risk assessment that acknowledges ambiguity in safety monitoring.” 

IBM is also releasing Granite Guardian 3.2 3B-A800M, which was created by fine-tuning the company’s mixture of experts (MoE) base model. 

All of the new Granite 3.2 models are available on Hugging Face under the Apache 2.0 license. Additionally, some of the models are accessible through IBM watsonx.ai, Ollama, Replicate, and LM Studio.